Applied-Software/ZLUDA
5
0
Fork 0
mirror of https://github.com/vosen/ZLUDA.git synced 2025-07-25 13:16:23 +03:00
Code Issues Packages Projects Releases Wiki Activity

Emit correct alignment for loads and stores (#429)

Browse Source
This commit is contained in:
Violet
2025-07-23 14:55:52 -07:00
committed by GitHub
parent a86ba3d642
commit 119b635b9d
113 changed files with 1117 additions and 1108 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
ptx/src/pass/llvm/emit.rs
View File

@ -532,8 +532,10 @@ impl<'a> MethodEmitContext<'a> {
let builder = self.builder;
let type_ = get_type(self.context, &data.typ)?;
let ptr = self.resolver.value(arguments.src)?;
self.resolver.with_result(arguments.dst, |dst| unsafe {
LLVMBuildLoad2(builder, type_, ptr, dst)
self.resolver.with_result(arguments.dst, |dst| {
let load = unsafe { LLVMBuildLoad2(builder, type_, ptr, dst) };
unsafe { LLVMSetAlignment(load, data.typ.layout().align() as u32) };
load
});
Ok(())
}
@ -739,7 +741,8 @@ impl<'a> MethodEmitContext<'a> {
if data.qualifier != ast::LdStQualifier::Weak {
todo!()
}
unsafe { LLVMBuildStore(self.builder, value, ptr) };
let store = unsafe { LLVMBuildStore(self.builder, value, ptr) };
unsafe { LLVMSetAlignment(store, data.typ.layout().align() as u32); }
Ok(())
}
@ -953,8 +956,14 @@ impl<'a> MethodEmitContext<'a> {
.iter()
.map(|(value, type_)| {
let value = self.resolver.value(*value)?;
let type_ = get_type(self.context, type_)?;
Ok(unsafe { LLVMBuildLoad2(self.builder, type_, value, LLVM_UNNAMED.as_ptr()) })
let lowered_type = get_type(self.context, type_)?;
let load = unsafe {
LLVMBuildLoad2(self.builder, lowered_type, value, LLVM_UNNAMED.as_ptr())
};
unsafe {
LLVMSetAlignment(load, type_.layout().align() as u32);
}
Ok(load)
})
.collect::<Result<Vec<_>, _>>()?;

6
ptx/src/test/ll/activemask.ll
View File

@ -9,11 +9,11 @@ define amdgpu_kernel void @activemask(ptr addrspace(4) byref(i64) %"29", ptr add
br label %"28"
"28": ; preds = %1
%"33" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"33", ptr addrspace(5) %"31", align 4
%"33" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"33", ptr addrspace(5) %"31", align 8
%"34" = call i32 @__zluda_ptx_impl_activemask()
store i32 %"34", ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(5) %"31", align 4
%"35" = load i64, ptr addrspace(5) %"31", align 8
%"36" = load i32, ptr addrspace(5) %"32", align 4
%"37" = inttoptr i64 %"35" to ptr
store i32 %"36", ptr %"37", align 4

26
ptx/src/test/ll/add.ll
View File

@ -9,22 +9,22 @@ define amdgpu_kernel void @add(ptr addrspace(4) byref(i64) %"32", ptr addrspace(
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr
%"40" = load i64, ptr %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"42" = add i64 %"43", 1
store i64 %"42", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"42", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"47" = inttoptr i64 %"44" to ptr
store i64 %"45", ptr %"47", align 4
store i64 %"45", ptr %"47", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

26
ptx/src/test/ll/add_non_coherent.ll
View File

@ -9,22 +9,22 @@ define amdgpu_kernel void @add_non_coherent(ptr addrspace(4) byref(i64) %"32", p
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr addrspace(1)
%"40" = load i64, ptr addrspace(1) %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr addrspace(1) %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"42" = add i64 %"43", 1
store i64 %"42", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"42", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"47" = inttoptr i64 %"44" to ptr addrspace(1)
store i64 %"45", ptr addrspace(1) %"47", align 4
store i64 %"45", ptr addrspace(1) %"47", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

16
ptx/src/test/ll/add_s32_sat.ll
View File

@ -11,15 +11,15 @@ define amdgpu_kernel void @add_s32_sat(ptr addrspace(4) byref(i64) %"37", ptr ad
br label %"36"
"36": ; preds = %1
%"45" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"45", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(4) %"38", align 4
store i64 %"46", ptr addrspace(5) %"40", align 4
%"48" = load i64, ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"45", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(4) %"38", align 8
store i64 %"46", ptr addrspace(5) %"40", align 8
%"48" = load i64, ptr addrspace(5) %"39", align 8
%"61" = inttoptr i64 %"48" to ptr
%"47" = load i32, ptr %"61", align 4
store i32 %"47", ptr addrspace(5) %"41", align 4
%"49" = load i64, ptr addrspace(5) %"39", align 4
%"49" = load i64, ptr addrspace(5) %"39", align 8
%"62" = inttoptr i64 %"49" to ptr
%"33" = getelementptr inbounds i8, ptr %"62", i64 4
%"50" = load i32, ptr %"33", align 4
@ -32,11 +32,11 @@ define amdgpu_kernel void @add_s32_sat(ptr addrspace(4) byref(i64) %"37", ptr ad
%"56" = load i32, ptr addrspace(5) %"42", align 4
%"54" = add i32 %"55", %"56"
store i32 %"54", ptr addrspace(5) %"44", align 4
%"57" = load i64, ptr addrspace(5) %"40", align 4
%"57" = load i64, ptr addrspace(5) %"40", align 8
%"58" = load i32, ptr addrspace(5) %"43", align 4
%"63" = inttoptr i64 %"57" to ptr
store i32 %"58", ptr %"63", align 4
%"59" = load i64, ptr addrspace(5) %"40", align 4
%"59" = load i64, ptr addrspace(5) %"40", align 8
%"64" = inttoptr i64 %"59" to ptr
%"35" = getelementptr inbounds i8, ptr %"64", i64 4
%"60" = load i32, ptr addrspace(5) %"44", align 4

26
ptx/src/test/ll/add_tuning.ll
View File

@ -9,22 +9,22 @@ define amdgpu_kernel void @add_tuning(ptr addrspace(4) byref(i64) %"32", ptr add
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr
%"40" = load i64, ptr %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"42" = add i64 %"43", 1
store i64 %"42", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"42", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"47" = inttoptr i64 %"44" to ptr
store i64 %"45", ptr %"47", align 4
store i64 %"45", ptr %"47", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

14
ptx/src/test/ll/and.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @and(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load i32, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @and(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
%"47" = load i32, ptr addrspace(5) %"38", align 4
%"52" = and i32 %"46", %"47"
store i32 %"52", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load i32, ptr addrspace(5) %"37", align 4
%"55" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"55", align 4

48
ptx/src/test/ll/assertfail.ll
View File

@ -17,46 +17,46 @@ define amdgpu_kernel void @assertfail(ptr addrspace(4) byref(i64) %"86", ptr add
br label %"84"
"84": ; preds = %1
%"92" = load i64, ptr addrspace(4) %"86", align 4
store i64 %"92", ptr addrspace(5) %"88", align 4
%"93" = load i64, ptr addrspace(4) %"87", align 4
store i64 %"93", ptr addrspace(5) %"89", align 4
%"92" = load i64, ptr addrspace(4) %"86", align 8
store i64 %"92", ptr addrspace(5) %"88", align 8
%"93" = load i64, ptr addrspace(4) %"87", align 8
store i64 %"93", ptr addrspace(5) %"89", align 8
store i32 0, ptr addrspace(5) %"94", align 4
%"97" = getelementptr inbounds i8, ptr addrspace(5) %"96", i64 0
%"98" = load i64, ptr addrspace(5) %"88", align 4
store i64 %"98", ptr addrspace(5) %"97", align 4
%"98" = load i64, ptr addrspace(5) %"88", align 8
store i64 %"98", ptr addrspace(5) %"97", align 8
%"100" = getelementptr inbounds i8, ptr addrspace(5) %"99", i64 0
%"101" = load i64, ptr addrspace(5) %"88", align 4
store i64 %"101", ptr addrspace(5) %"100", align 4
%"101" = load i64, ptr addrspace(5) %"88", align 8
store i64 %"101", ptr addrspace(5) %"100", align 8
%"103" = getelementptr inbounds i8, ptr addrspace(5) %"102", i64 0
%"104" = load i32, ptr addrspace(5) %"94", align 4
store i32 %"104", ptr addrspace(5) %"103", align 4
%"106" = getelementptr inbounds i8, ptr addrspace(5) %"105", i64 0
%"107" = load i64, ptr addrspace(5) %"88", align 4
store i64 %"107", ptr addrspace(5) %"106", align 4
%"107" = load i64, ptr addrspace(5) %"88", align 8
store i64 %"107", ptr addrspace(5) %"106", align 8
%"109" = getelementptr inbounds i8, ptr addrspace(5) %"108", i64 0
%"110" = load i64, ptr addrspace(5) %"88", align 4
store i64 %"110", ptr addrspace(5) %"109", align 4
%"74" = load i64, ptr addrspace(5) %"96", align 4
%"75" = load i64, ptr addrspace(5) %"99", align 4
%"110" = load i64, ptr addrspace(5) %"88", align 8
store i64 %"110", ptr addrspace(5) %"109", align 8
%"74" = load i64, ptr addrspace(5) %"96", align 8
%"75" = load i64, ptr addrspace(5) %"99", align 8
%"76" = load i32, ptr addrspace(5) %"102", align 4
%"77" = load i64, ptr addrspace(5) %"105", align 4
%"78" = load i64, ptr addrspace(5) %"108", align 4
%"77" = load i64, ptr addrspace(5) %"105", align 8
%"78" = load i64, ptr addrspace(5) %"108", align 8
call void @__zluda_ptx_impl___assertfail(i64 %"74", i64 %"75", i32 %"76", i64 %"77", i64 %"78")
br label %"85"
"85": ; preds = %"84"
%"112" = load i64, ptr addrspace(5) %"88", align 4
%"112" = load i64, ptr addrspace(5) %"88", align 8
%"122" = inttoptr i64 %"112" to ptr
%"111" = load i64, ptr %"122", align 4
store i64 %"111", ptr addrspace(5) %"90", align 4
%"114" = load i64, ptr addrspace(5) %"90", align 4
%"111" = load i64, ptr %"122", align 8
store i64 %"111", ptr addrspace(5) %"90", align 8
%"114" = load i64, ptr addrspace(5) %"90", align 8
%"113" = add i64 %"114", 1
store i64 %"113", ptr addrspace(5) %"91", align 4
%"115" = load i64, ptr addrspace(5) %"89", align 4
%"116" = load i64, ptr addrspace(5) %"91", align 4
store i64 %"113", ptr addrspace(5) %"91", align 8
%"115" = load i64, ptr addrspace(5) %"89", align 8
%"116" = load i64, ptr addrspace(5) %"91", align 8
%"123" = inttoptr i64 %"115" to ptr
store i64 %"116", ptr %"123", align 4
store i64 %"116", ptr %"123", align 8
ret void
}

16
ptx/src/test/ll/atom_add.ll
View File

@ -11,15 +11,15 @@ define amdgpu_kernel void @atom_add(ptr addrspace(4) byref(i64) %"36", ptr addrs
br label %"35"
"35": ; preds = %1
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"38", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"43" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"43", ptr addrspace(5) %"39", align 8
%"45" = load i64, ptr addrspace(5) %"38", align 8
%"56" = inttoptr i64 %"45" to ptr
%"44" = load i32, ptr %"56", align 4
store i32 %"44", ptr addrspace(5) %"40", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 8
%"57" = inttoptr i64 %"46" to ptr
%"32" = getelementptr inbounds i8, ptr %"57", i64 4
%"47" = load i32, ptr %"32", align 4
@ -31,11 +31,11 @@ define amdgpu_kernel void @atom_add(ptr addrspace(4) byref(i64) %"36", ptr addrs
store i32 %2, ptr addrspace(5) %"40", align 4
%"51" = load i32, ptr addrspace(3) @shared_mem, align 4
store i32 %"51", ptr addrspace(5) %"41", align 4
%"52" = load i64, ptr addrspace(5) %"39", align 4
%"52" = load i64, ptr addrspace(5) %"39", align 8
%"53" = load i32, ptr addrspace(5) %"40", align 4
%"61" = inttoptr i64 %"52" to ptr
store i32 %"53", ptr %"61", align 4
%"54" = load i64, ptr addrspace(5) %"39", align 4
%"54" = load i64, ptr addrspace(5) %"39", align 8
%"62" = inttoptr i64 %"54" to ptr
%"34" = getelementptr inbounds i8, ptr %"62", i64 4
%"55" = load i32, ptr addrspace(5) %"41", align 4

16
ptx/src/test/ll/atom_add_float.ll
View File

@ -11,15 +11,15 @@ define amdgpu_kernel void @atom_add_float(ptr addrspace(4) byref(i64) %"36", ptr
br label %"35"
"35": ; preds = %1
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"38", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"43" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"43", ptr addrspace(5) %"39", align 8
%"45" = load i64, ptr addrspace(5) %"38", align 8
%"56" = inttoptr i64 %"45" to ptr
%"44" = load float, ptr %"56", align 4
store float %"44", ptr addrspace(5) %"40", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 8
%"57" = inttoptr i64 %"46" to ptr
%"32" = getelementptr inbounds i8, ptr %"57", i64 4
%"47" = load float, ptr %"32", align 4
@ -31,11 +31,11 @@ define amdgpu_kernel void @atom_add_float(ptr addrspace(4) byref(i64) %"36", ptr
store float %2, ptr addrspace(5) %"40", align 4
%"51" = load float, ptr addrspace(3) @shared_mem, align 4
store float %"51", ptr addrspace(5) %"41", align 4
%"52" = load i64, ptr addrspace(5) %"39", align 4
%"52" = load i64, ptr addrspace(5) %"39", align 8
%"53" = load float, ptr addrspace(5) %"40", align 4
%"61" = inttoptr i64 %"52" to ptr
store float %"53", ptr %"61", align 4
%"54" = load i64, ptr addrspace(5) %"39", align 4
%"54" = load i64, ptr addrspace(5) %"39", align 8
%"62" = inttoptr i64 %"54" to ptr
%"34" = getelementptr inbounds i8, ptr %"62", i64 4
%"55" = load float, ptr addrspace(5) %"41", align 4

18
ptx/src/test/ll/atom_cas.ll
View File

@ -9,31 +9,31 @@ define amdgpu_kernel void @atom_cas(ptr addrspace(4) byref(i64) %"38", ptr addrs
br label %"37"
"37": ; preds = %1
%"44" = load i64, ptr addrspace(4) %"38", align 4
store i64 %"44", ptr addrspace(5) %"40", align 4
%"45" = load i64, ptr addrspace(4) %"39", align 4
store i64 %"45", ptr addrspace(5) %"41", align 4
%"47" = load i64, ptr addrspace(5) %"40", align 4
%"44" = load i64, ptr addrspace(4) %"38", align 8
store i64 %"44", ptr addrspace(5) %"40", align 8
%"45" = load i64, ptr addrspace(4) %"39", align 8
store i64 %"45", ptr addrspace(5) %"41", align 8
%"47" = load i64, ptr addrspace(5) %"40", align 8
%"57" = inttoptr i64 %"47" to ptr
%"46" = load i32, ptr %"57", align 4
store i32 %"46", ptr addrspace(5) %"42", align 4
%"48" = load i64, ptr addrspace(5) %"40", align 4
%"48" = load i64, ptr addrspace(5) %"40", align 8
%"58" = inttoptr i64 %"48" to ptr
%"31" = getelementptr inbounds i8, ptr %"58", i64 4
%"50" = load i32, ptr addrspace(5) %"42", align 4
%2 = cmpxchg ptr %"31", i32 %"50", i32 100 syncscope("agent-one-as") monotonic monotonic, align 4
%"59" = extractvalue { i32, i1 } %2, 0
store i32 %"59", ptr addrspace(5) %"42", align 4
%"51" = load i64, ptr addrspace(5) %"40", align 4
%"51" = load i64, ptr addrspace(5) %"40", align 8
%"61" = inttoptr i64 %"51" to ptr
%"34" = getelementptr inbounds i8, ptr %"61", i64 4
%"52" = load i32, ptr %"34", align 4
store i32 %"52", ptr addrspace(5) %"43", align 4
%"53" = load i64, ptr addrspace(5) %"41", align 4
%"53" = load i64, ptr addrspace(5) %"41", align 8
%"54" = load i32, ptr addrspace(5) %"42", align 4
%"62" = inttoptr i64 %"53" to ptr
store i32 %"54", ptr %"62", align 4
%"55" = load i64, ptr addrspace(5) %"41", align 4
%"55" = load i64, ptr addrspace(5) %"41", align 8
%"63" = inttoptr i64 %"55" to ptr
%"36" = getelementptr inbounds i8, ptr %"63", i64 4
%"56" = load i32, ptr addrspace(5) %"43", align 4

20
ptx/src/test/ll/atom_inc.ll
View File

@ -10,32 +10,32 @@ define amdgpu_kernel void @atom_inc(ptr addrspace(4) byref(i64) %"38", ptr addrs
br label %"37"
"37": ; preds = %1
%"45" = load i64, ptr addrspace(4) %"38", align 4
store i64 %"45", ptr addrspace(5) %"40", align 4
%"46" = load i64, ptr addrspace(4) %"39", align 4
store i64 %"46", ptr addrspace(5) %"41", align 4
%"48" = load i64, ptr addrspace(5) %"40", align 4
%"45" = load i64, ptr addrspace(4) %"38", align 8
store i64 %"45", ptr addrspace(5) %"40", align 8
%"46" = load i64, ptr addrspace(4) %"39", align 8
store i64 %"46", ptr addrspace(5) %"41", align 8
%"48" = load i64, ptr addrspace(5) %"40", align 8
%"59" = inttoptr i64 %"48" to ptr
%2 = atomicrmw uinc_wrap ptr %"59", i32 101 syncscope("agent-one-as") monotonic, align 4
store i32 %2, ptr addrspace(5) %"42", align 4
%"50" = load i64, ptr addrspace(5) %"40", align 4
%"50" = load i64, ptr addrspace(5) %"40", align 8
%"60" = inttoptr i64 %"50" to ptr addrspace(1)
%3 = atomicrmw uinc_wrap ptr addrspace(1) %"60", i32 101 syncscope("agent-one-as") monotonic, align 4
store i32 %3, ptr addrspace(5) %"43", align 4
%"52" = load i64, ptr addrspace(5) %"40", align 4
%"52" = load i64, ptr addrspace(5) %"40", align 8
%"61" = inttoptr i64 %"52" to ptr
%"51" = load i32, ptr %"61", align 4
store i32 %"51", ptr addrspace(5) %"44", align 4
%"53" = load i64, ptr addrspace(5) %"41", align 4
%"53" = load i64, ptr addrspace(5) %"41", align 8
%"54" = load i32, ptr addrspace(5) %"42", align 4
%"62" = inttoptr i64 %"53" to ptr
store i32 %"54", ptr %"62", align 4
%"55" = load i64, ptr addrspace(5) %"41", align 4
%"55" = load i64, ptr addrspace(5) %"41", align 8
%"63" = inttoptr i64 %"55" to ptr
%"34" = getelementptr inbounds i8, ptr %"63", i64 4
%"56" = load i32, ptr addrspace(5) %"43", align 4
store i32 %"56", ptr %"34", align 4
%"57" = load i64, ptr addrspace(5) %"41", align 4
%"57" = load i64, ptr addrspace(5) %"41", align 8
%"64" = inttoptr i64 %"57" to ptr
%"36" = getelementptr inbounds i8, ptr %"64", i64 8
%"58" = load i32, ptr addrspace(5) %"44", align 4

20
ptx/src/test/ll/b64tof64.ll
View File

@ -11,20 +11,20 @@ define amdgpu_kernel void @b64tof64(ptr addrspace(4) byref(i64) %"31", ptr addrs
"30": ; preds = %1
%"37" = load double, ptr addrspace(4) %"31", align 8
store double %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"35", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"35", align 8
%"40" = load double, ptr addrspace(5) %"33", align 8
%"46" = bitcast double %"40" to i64
store i64 %"46", ptr addrspace(5) %"34", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
store i64 %"46", ptr addrspace(5) %"34", align 8
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"47" = inttoptr i64 %"42" to ptr
%"41" = load i64, ptr %"47", align 4
store i64 %"41", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr %"47", align 8
store i64 %"41", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"48" = inttoptr i64 %"43" to ptr
store i64 %"44", ptr %"48", align 4
store i64 %"44", ptr %"48", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

18
ptx/src/test/ll/bar_red_and_pred.ll
View File

@ -18,8 +18,8 @@ define amdgpu_kernel void @bar_red_and_pred(ptr addrspace(4) byref(i64) %"73", p
br label %"70"
"70": ; preds = %1
%"82" = load i64, ptr addrspace(4) %"74", align 4
store i64 %"82", ptr addrspace(5) %"75", align 4
%"82" = load i64, ptr addrspace(4) %"74", align 8
store i64 %"82", ptr addrspace(5) %"75", align 8
%"44" = call i32 @__zluda_ptx_impl_sreg_tid(i8 0)
br label %"71"
@ -102,15 +102,15 @@ define amdgpu_kernel void @bar_red_and_pred(ptr addrspace(4) byref(i64) %"73", p
"26": ; preds = %"25", %"24"
%"118" = load i32, ptr addrspace(5) %"77", align 4
%"117" = zext i32 %"118" to i64
store i64 %"117", ptr addrspace(5) %"76", align 4
%"120" = load i64, ptr addrspace(5) %"76", align 4
store i64 %"117", ptr addrspace(5) %"76", align 8
%"120" = load i64, ptr addrspace(5) %"76", align 8
%"119" = mul i64 %"120", 4
store i64 %"119", ptr addrspace(5) %"76", align 4
%"122" = load i64, ptr addrspace(5) %"75", align 4
%"123" = load i64, ptr addrspace(5) %"76", align 4
store i64 %"119", ptr addrspace(5) %"76", align 8
%"122" = load i64, ptr addrspace(5) %"75", align 8
%"123" = load i64, ptr addrspace(5) %"76", align 8
%"121" = add i64 %"122", %"123"
store i64 %"121", ptr addrspace(5) %"75", align 4
%"124" = load i64, ptr addrspace(5) %"75", align 4
store i64 %"121", ptr addrspace(5) %"75", align 8
%"124" = load i64, ptr addrspace(5) %"75", align 8
%"125" = load i32, ptr addrspace(5) %"81", align 4
%"126" = inttoptr i64 %"124" to ptr
store i32 %"125", ptr %"126", align 4

16
ptx/src/test/ll/bfe.ll
View File

@ -12,20 +12,20 @@ define amdgpu_kernel void @bfe(ptr addrspace(4) byref(i64) %"36", ptr addrspace(
br label %"35"
"35": ; preds = %1
%"43" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"43", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"44", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"43", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"44", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(5) %"38", align 8
%"57" = inttoptr i64 %"46" to ptr
%"45" = load i32, ptr %"57", align 4
store i32 %"45", ptr addrspace(5) %"40", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 8
%"58" = inttoptr i64 %"47" to ptr
%"32" = getelementptr inbounds i8, ptr %"58", i64 4
%"48" = load i32, ptr %"32", align 4
store i32 %"48", ptr addrspace(5) %"41", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 8
%"59" = inttoptr i64 %"49" to ptr
%"34" = getelementptr inbounds i8, ptr %"59", i64 8
%"50" = load i32, ptr %"34", align 4
@ -35,7 +35,7 @@ define amdgpu_kernel void @bfe(ptr addrspace(4) byref(i64) %"36", ptr addrspace(
%"54" = load i32, ptr addrspace(5) %"42", align 4
%"51" = call i32 @__zluda_ptx_impl_bfe_u32(i32 %"52", i32 %"53", i32 %"54")
store i32 %"51", ptr addrspace(5) %"40", align 4
%"55" = load i64, ptr addrspace(5) %"39", align 4
%"55" = load i64, ptr addrspace(5) %"39", align 8
%"56" = load i32, ptr addrspace(5) %"40", align 4
%"60" = inttoptr i64 %"55" to ptr
store i32 %"56", ptr %"60", align 4

18
ptx/src/test/ll/bfi.ll
View File

@ -13,25 +13,25 @@ define amdgpu_kernel void @bfi(ptr addrspace(4) byref(i64) %"39", ptr addrspace(
br label %"38"
"38": ; preds = %1
%"47" = load i64, ptr addrspace(4) %"39", align 4
store i64 %"47", ptr addrspace(5) %"41", align 4
%"48" = load i64, ptr addrspace(4) %"40", align 4
store i64 %"48", ptr addrspace(5) %"42", align 4
%"50" = load i64, ptr addrspace(5) %"41", align 4
%"47" = load i64, ptr addrspace(4) %"39", align 8
store i64 %"47", ptr addrspace(5) %"41", align 8
%"48" = load i64, ptr addrspace(4) %"40", align 8
store i64 %"48", ptr addrspace(5) %"42", align 8
%"50" = load i64, ptr addrspace(5) %"41", align 8
%"64" = inttoptr i64 %"50" to ptr
%"49" = load i32, ptr %"64", align 4
store i32 %"49", ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr addrspace(5) %"41", align 4
%"51" = load i64, ptr addrspace(5) %"41", align 8
%"65" = inttoptr i64 %"51" to ptr
%"33" = getelementptr inbounds i8, ptr %"65", i64 4
%"52" = load i32, ptr %"33", align 4
store i32 %"52", ptr addrspace(5) %"44", align 4
%"53" = load i64, ptr addrspace(5) %"41", align 4
%"53" = load i64, ptr addrspace(5) %"41", align 8
%"66" = inttoptr i64 %"53" to ptr
%"35" = getelementptr inbounds i8, ptr %"66", i64 8
%"54" = load i32, ptr %"35", align 4
store i32 %"54", ptr addrspace(5) %"45", align 4
%"55" = load i64, ptr addrspace(5) %"41", align 4
%"55" = load i64, ptr addrspace(5) %"41", align 8
%"67" = inttoptr i64 %"55" to ptr
%"37" = getelementptr inbounds i8, ptr %"67", i64 12
%"56" = load i32, ptr %"37", align 4
@ -42,7 +42,7 @@ define amdgpu_kernel void @bfi(ptr addrspace(4) byref(i64) %"39", ptr addrspace(
%"61" = load i32, ptr addrspace(5) %"46", align 4
%"68" = call i32 @__zluda_ptx_impl_bfi_b32(i32 %"58", i32 %"59", i32 %"60", i32 %"61")
store i32 %"68", ptr addrspace(5) %"43", align 4
%"62" = load i64, ptr addrspace(5) %"42", align 4
%"62" = load i64, ptr addrspace(5) %"42", align 8
%"63" = load i32, ptr addrspace(5) %"43", align 4
%"71" = inttoptr i64 %"62" to ptr
store i32 %"63", ptr %"71", align 4

30
ptx/src/test/ll/block.ll
View File

@ -10,25 +10,25 @@ define amdgpu_kernel void @block(ptr addrspace(4) byref(i64) %"34", ptr addrspac
br label %"33"
"33": ; preds = %1
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"51" = inttoptr i64 %"43" to ptr
%"42" = load i64, ptr %"51", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"45" = load i64, ptr addrspace(5) %"38", align 4
%"42" = load i64, ptr %"51", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"45" = load i64, ptr addrspace(5) %"38", align 8
%"44" = add i64 %"45", 1
store i64 %"44", ptr addrspace(5) %"39", align 4
%"48" = load i64, ptr addrspace(5) %"46", align 4
store i64 %"44", ptr addrspace(5) %"39", align 8
%"48" = load i64, ptr addrspace(5) %"46", align 8
%"47" = add i64 %"48", 1
store i64 %"47", ptr addrspace(5) %"46", align 4
%"49" = load i64, ptr addrspace(5) %"37", align 4
%"50" = load i64, ptr addrspace(5) %"39", align 4
store i64 %"47", ptr addrspace(5) %"46", align 8
%"49" = load i64, ptr addrspace(5) %"37", align 8
%"50" = load i64, ptr addrspace(5) %"39", align 8
%"52" = inttoptr i64 %"49" to ptr
store i64 %"50", ptr %"52", align 4
store i64 %"50", ptr %"52", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

26
ptx/src/test/ll/bra.ll
View File

@ -9,28 +9,28 @@ define amdgpu_kernel void @bra(ptr addrspace(4) byref(i64) %"36", ptr addrspace(
br label %"35"
"35": ; preds = %1
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"38", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"43" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"43", ptr addrspace(5) %"39", align 8
%"45" = load i64, ptr addrspace(5) %"38", align 8
%"50" = inttoptr i64 %"45" to ptr
%"44" = load i64, ptr %"50", align 4
store i64 %"44", ptr addrspace(5) %"40", align 4
%"44" = load i64, ptr %"50", align 8
store i64 %"44", ptr addrspace(5) %"40", align 8
br label %"10"
"10": ; preds = %"35"
%"47" = load i64, ptr addrspace(5) %"40", align 4
%"47" = load i64, ptr addrspace(5) %"40", align 8
%"46" = add i64 %"47", 1
store i64 %"46", ptr addrspace(5) %"41", align 4
store i64 %"46", ptr addrspace(5) %"41", align 8
br label %"12"
"12": ; preds = %"10"
%"48" = load i64, ptr addrspace(5) %"39", align 4
%"49" = load i64, ptr addrspace(5) %"41", align 4
%"48" = load i64, ptr addrspace(5) %"39", align 8
%"49" = load i64, ptr addrspace(5) %"41", align 8
%"51" = inttoptr i64 %"48" to ptr
store i64 %"49", ptr %"51", align 4
store i64 %"49", ptr %"51", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

12
ptx/src/test/ll/brev.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @brev(ptr addrspace(4) byref(i64) %"30", ptr addrspace
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load i32, ptr %"43", align 4
store i32 %"37", ptr addrspace(5) %"34", align 4
%"40" = load i32, ptr addrspace(5) %"34", align 4
%"39" = call i32 @llvm.bitreverse.i32(i32 %"40")
store i32 %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load i32, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store i32 %"42", ptr %"44", align 4

54
ptx/src/test/ll/call.ll
View File

@ -9,17 +9,17 @@ define i64 @incr(i64 %"43") #0 {
br label %"46"
"46": ; preds = %1
store i64 %"43", ptr addrspace(5) %"65", align 4
%"67" = load i64, ptr addrspace(5) %"65", align 4
store i64 %"67", ptr addrspace(5) %"66", align 4
%"69" = load i64, ptr addrspace(5) %"66", align 4
store i64 %"43", ptr addrspace(5) %"65", align 8
%"67" = load i64, ptr addrspace(5) %"65", align 8
store i64 %"67", ptr addrspace(5) %"66", align 8
%"69" = load i64, ptr addrspace(5) %"66", align 8
%"68" = add i64 %"69", 1
store i64 %"68", ptr addrspace(5) %"66", align 4
%"70" = load i64, ptr addrspace(5) %"66", align 4
store i64 %"70", ptr addrspace(5) %"64", align 4
%"71" = load i64, ptr addrspace(5) %"64", align 4
store i64 %"71", ptr addrspace(5) %"63", align 4
%2 = load i64, ptr addrspace(5) %"63", align 4
store i64 %"68", ptr addrspace(5) %"66", align 8
%"70" = load i64, ptr addrspace(5) %"66", align 8
store i64 %"70", ptr addrspace(5) %"64", align 8
%"71" = load i64, ptr addrspace(5) %"64", align 8
store i64 %"71", ptr addrspace(5) %"63", align 8
%2 = load i64, ptr addrspace(5) %"63", align 8
ret i64 %2
}
@ -35,30 +35,30 @@ define amdgpu_kernel void @call(ptr addrspace(4) byref(i64) %"48", ptr addrspace
br label %"44"
"44": ; preds = %1
%"53" = load i64, ptr addrspace(4) %"48", align 4
store i64 %"53", ptr addrspace(5) %"50", align 4
%"54" = load i64, ptr addrspace(4) %"49", align 4
store i64 %"54", ptr addrspace(5) %"51", align 4
%"56" = load i64, ptr addrspace(5) %"50", align 4
%"53" = load i64, ptr addrspace(4) %"48", align 8
store i64 %"53", ptr addrspace(5) %"50", align 8
%"54" = load i64, ptr addrspace(4) %"49", align 8
store i64 %"54", ptr addrspace(5) %"51", align 8
%"56" = load i64, ptr addrspace(5) %"50", align 8
%"72" = inttoptr i64 %"56" to ptr addrspace(1)
%"55" = load i64, ptr addrspace(1) %"72", align 4
store i64 %"55", ptr addrspace(5) %"52", align 4
%"59" = load i64, ptr addrspace(5) %"52", align 4
store i64 %"59", ptr addrspace(5) %"57", align 4
%"40" = load i64, ptr addrspace(5) %"57", align 4
%"55" = load i64, ptr addrspace(1) %"72", align 8
store i64 %"55", ptr addrspace(5) %"52", align 8
%"59" = load i64, ptr addrspace(5) %"52", align 8
store i64 %"59", ptr addrspace(5) %"57", align 8
%"40" = load i64, ptr addrspace(5) %"57", align 8
%"41" = call i64 @incr(i64 %"40")
br label %"45"
"45": ; preds = %"44"
store i64 %"41", ptr addrspace(5) %"58", align 4
%"60" = load i64, ptr addrspace(5) %"58", align 4
store i64 %"60", ptr addrspace(5) %"52", align 4
%"61" = load i64, ptr addrspace(5) %"51", align 4
%"62" = load i64, ptr addrspace(5) %"52", align 4
store i64 %"41", ptr addrspace(5) %"58", align 8
%"60" = load i64, ptr addrspace(5) %"58", align 8
store i64 %"60", ptr addrspace(5) %"52", align 8
%"61" = load i64, ptr addrspace(5) %"51", align 8
%"62" = load i64, ptr addrspace(5) %"52", align 8
%"75" = inttoptr i64 %"61" to ptr addrspace(1)
store i64 %"62", ptr addrspace(1) %"75", align 4
store i64 %"62", ptr addrspace(1) %"75", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="dynamic" "denormal-fp-math-f32"="dynamic" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #1 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #1 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

20
ptx/src/test/ll/call_rnd.ll
View File

@ -86,25 +86,25 @@ define amdgpu_kernel void @call_rnd(ptr addrspace(4) byref(i64) %"92", ptr addrs
"84": ; preds = %1
call void @llvm.amdgcn.s.setreg(i32 6145, i32 1)
%"108" = load i64, ptr addrspace(4) %"92", align 4
store i64 %"108", ptr addrspace(5) %"94", align 4
%"109" = load i64, ptr addrspace(4) %"93", align 4
store i64 %"109", ptr addrspace(5) %"95", align 4
%"111" = load i64, ptr addrspace(5) %"94", align 4
%"108" = load i64, ptr addrspace(4) %"92", align 8
store i64 %"108", ptr addrspace(5) %"94", align 8
%"109" = load i64, ptr addrspace(4) %"93", align 8
store i64 %"109", ptr addrspace(5) %"95", align 8
%"111" = load i64, ptr addrspace(5) %"94", align 8
%"154" = inttoptr i64 %"111" to ptr
%"110" = load float, ptr %"154", align 4
store float %"110", ptr addrspace(5) %"96", align 4
%"112" = load i64, ptr addrspace(5) %"94", align 4
%"112" = load i64, ptr addrspace(5) %"94", align 8
%"155" = inttoptr i64 %"112" to ptr
%"59" = getelementptr inbounds i8, ptr %"155", i64 4
%"113" = load float, ptr %"59", align 4
store float %"113", ptr addrspace(5) %"97", align 4
%"114" = load i64, ptr addrspace(5) %"94", align 4
%"114" = load i64, ptr addrspace(5) %"94", align 8
%"156" = inttoptr i64 %"114" to ptr
%"61" = getelementptr inbounds i8, ptr %"156", i64 8
%"115" = load float, ptr %"61", align 4
store float %"115", ptr addrspace(5) %"98", align 4
%"116" = load i64, ptr addrspace(5) %"94", align 4
%"116" = load i64, ptr addrspace(5) %"94", align 8
%"157" = inttoptr i64 %"116" to ptr
%"63" = getelementptr inbounds i8, ptr %"157", i64 12
%"117" = load float, ptr %"63", align 4
@ -122,7 +122,7 @@ define amdgpu_kernel void @call_rnd(ptr addrspace(4) byref(i64) %"92", ptr addrs
store float %"74", ptr addrspace(5) %"104", align 4
%"120" = load float, ptr addrspace(5) %"104", align 4
store float %"120", ptr addrspace(5) %"100", align 4
%"121" = load i64, ptr addrspace(5) %"95", align 4
%"121" = load i64, ptr addrspace(5) %"95", align 8
%"122" = load float, ptr addrspace(5) %"100", align 4
%"158" = inttoptr i64 %"121" to ptr
store float %"122", ptr %"158", align 4
@ -139,7 +139,7 @@ define amdgpu_kernel void @call_rnd(ptr addrspace(4) byref(i64) %"92", ptr addrs
store float %"77", ptr addrspace(5) %"107", align 4
%"125" = load float, ptr addrspace(5) %"107", align 4
store float %"125", ptr addrspace(5) %"101", align 4
%"126" = load i64, ptr addrspace(5) %"95", align 4
%"126" = load i64, ptr addrspace(5) %"95", align 8
%"159" = inttoptr i64 %"126" to ptr
%"65" = getelementptr inbounds i8, ptr %"159", i64 4
%"127" = load float, ptr addrspace(5) %"101", align 4

12
ptx/src/test/ll/clz.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @clz(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load i32, ptr %"43", align 4
store i32 %"37", ptr addrspace(5) %"34", align 4
%"40" = load i32, ptr addrspace(5) %"34", align 4
%"44" = call i32 @llvm.ctlz.i32(i32 %"40", i1 false)
store i32 %"44", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load i32, ptr addrspace(5) %"34", align 4
%"45" = inttoptr i64 %"41" to ptr
store i32 %"42", ptr %"45", align 4

18
ptx/src/test/ll/const.ll
View File

@ -13,10 +13,10 @@ define amdgpu_kernel void @const(ptr addrspace(4) byref(i64) %"46", ptr addrspac
br label %"45"
"45": ; preds = %1
%"54" = load i64, ptr addrspace(4) %"46", align 4
store i64 %"54", ptr addrspace(5) %"48", align 4
%"55" = load i64, ptr addrspace(4) %"47", align 4
store i64 %"55", ptr addrspace(5) %"49", align 4
%"54" = load i64, ptr addrspace(4) %"46", align 8
store i64 %"54", ptr addrspace(5) %"48", align 8
%"55" = load i64, ptr addrspace(4) %"47", align 8
store i64 %"55", ptr addrspace(5) %"49", align 8
%"56" = load i16, ptr addrspace(4) @constparams, align 2
store i16 %"56", ptr addrspace(5) %"50", align 2
%"57" = load i16, ptr addrspace(4) getelementptr inbounds (i8, ptr addrspace(4) @constparams, i64 2), align 2
@ -25,21 +25,21 @@ define amdgpu_kernel void @const(ptr addrspace(4) byref(i64) %"46", ptr addrspac
store i16 %"58", ptr addrspace(5) %"52", align 2
%"59" = load i16, ptr addrspace(4) getelementptr inbounds (i8, ptr addrspace(4) @constparams, i64 6), align 2
store i16 %"59", ptr addrspace(5) %"53", align 2
%"60" = load i64, ptr addrspace(5) %"49", align 4
%"60" = load i64, ptr addrspace(5) %"49", align 8
%"61" = load i16, ptr addrspace(5) %"50", align 2
%"72" = inttoptr i64 %"60" to ptr
store i16 %"61", ptr %"72", align 2
%"62" = load i64, ptr addrspace(5) %"49", align 4
%"62" = load i64, ptr addrspace(5) %"49", align 8
%"74" = inttoptr i64 %"62" to ptr
%"40" = getelementptr inbounds i8, ptr %"74", i64 2
%"63" = load i16, ptr addrspace(5) %"51", align 2
store i16 %"63", ptr %"40", align 2
%"64" = load i64, ptr addrspace(5) %"49", align 4
%"64" = load i64, ptr addrspace(5) %"49", align 8
%"76" = inttoptr i64 %"64" to ptr
%"42" = getelementptr inbounds i8, ptr %"76", i64 4
%"65" = load i16, ptr addrspace(5) %"52", align 2
store i16 %"65", ptr %"42", align 2
%"66" = load i64, ptr addrspace(5) %"49", align 4
%"66" = load i64, ptr addrspace(5) %"49", align 8
%"78" = inttoptr i64 %"66" to ptr
%"44" = getelementptr inbounds i8, ptr %"78", i64 6
%"67" = load i16, ptr addrspace(5) %"53", align 2
@ -47,4 +47,4 @@ define amdgpu_kernel void @const(ptr addrspace(4) byref(i64) %"46", ptr addrspac
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

12
ptx/src/test/ll/constant_f32.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @constant_f32(ptr addrspace(4) byref(i64) %"31", ptr a
br label %"30"
"30": ; preds = %1
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"44" = inttoptr i64 %"39" to ptr
%"38" = load float, ptr %"44", align 4
store float %"38", ptr addrspace(5) %"35", align 4
%"41" = load float, ptr addrspace(5) %"35", align 4
%"40" = fmul float %"41", 5.000000e-01
store float %"40", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"43" = load float, ptr addrspace(5) %"35", align 4
%"45" = inttoptr i64 %"42" to ptr
store float %"43", ptr %"45", align 4

12
ptx/src/test/ll/constant_negative.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @constant_negative(ptr addrspace(4) byref(i64) %"31",
br label %"30"
"30": ; preds = %1
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"44" = inttoptr i64 %"39" to ptr
%"38" = load i32, ptr %"44", align 4
store i32 %"38", ptr addrspace(5) %"35", align 4
%"41" = load i32, ptr addrspace(5) %"35", align 4
%"40" = mul i32 %"41", -1
store i32 %"40", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"43" = load i32, ptr addrspace(5) %"35", align 4
%"45" = inttoptr i64 %"42" to ptr
store i32 %"43", ptr %"45", align 4

12
ptx/src/test/ll/cos.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @cos(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load float, ptr %"43", align 4
store float %"37", ptr addrspace(5) %"34", align 4
%"40" = load float, ptr addrspace(5) %"34", align 4
%"39" = call afn float @llvm.cos.f32(float %"40")
store float %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load float, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store float %"42", ptr %"44", align 4

14
ptx/src/test/ll/cvt_f64_f32.ll
View File

@ -9,22 +9,22 @@ define amdgpu_kernel void @cvt_f64_f32(ptr addrspace(4) byref(i64) %"31", ptr ad
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"45" = inttoptr i64 %"40" to ptr addrspace(1)
%"39" = load float, ptr addrspace(1) %"45", align 4
store float %"39", ptr addrspace(5) %"35", align 4
%"42" = load float, ptr addrspace(5) %"35", align 4
%"41" = fpext float %"42" to double
store double %"41", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"34", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 8
%"44" = load double, ptr addrspace(5) %"36", align 8
%"46" = inttoptr i64 %"43" to ptr
store double %"44", ptr %"46", align 8
ret void
}
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }

18
ptx/src/test/ll/cvt_rni.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @cvt_rni(ptr addrspace(4) byref(i64) %"35", ptr addrsp
br label %"34"
"34": ; preds = %1
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(5) %"37", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(5) %"37", align 8
%"55" = inttoptr i64 %"44" to ptr
%"43" = load float, ptr %"55", align 4
store float %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"56" = inttoptr i64 %"45" to ptr
%"31" = getelementptr inbounds i8, ptr %"56", i64 4
%"46" = load float, ptr %"31", align 4
@ -30,11 +30,11 @@ define amdgpu_kernel void @cvt_rni(ptr addrspace(4) byref(i64) %"35", ptr addrsp
%3 = call float @llvm.roundeven.f32(float %"50")
%"49" = freeze float %3
store float %"49", ptr addrspace(5) %"40", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 8
%"52" = load float, ptr addrspace(5) %"39", align 4
%"57" = inttoptr i64 %"51" to ptr
store float %"52", ptr %"57", align 4
%"53" = load i64, ptr addrspace(5) %"38", align 4
%"53" = load i64, ptr addrspace(5) %"38", align 8
%"58" = inttoptr i64 %"53" to ptr
%"33" = getelementptr inbounds i8, ptr %"58", i64 4
%"54" = load float, ptr addrspace(5) %"40", align 4
@ -46,4 +46,4 @@ define amdgpu_kernel void @cvt_rni(ptr addrspace(4) byref(i64) %"35", ptr addrsp
declare float @llvm.roundeven.f32(float) #1
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="ieee" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }

12
ptx/src/test/ll/cvt_rni_u16_f32.ll
View File

@ -9,11 +9,11 @@ define amdgpu_kernel void @cvt_rni_u16_f32(ptr addrspace(4) byref(i64) %"31", pt
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"45" = inttoptr i64 %"40" to ptr addrspace(1)
%"39" = load float, ptr addrspace(1) %"45", align 4
store float %"39", ptr addrspace(5) %"35", align 4
@ -21,7 +21,7 @@ define amdgpu_kernel void @cvt_rni_u16_f32(ptr addrspace(4) byref(i64) %"31", pt
%2 = call float @llvm.roundeven.f32(float %"42")
%"41" = call i16 @llvm.fptoui.sat.i16.f32(float %2)
store i16 %"41", ptr addrspace(5) %"36", align 2
%"43" = load i64, ptr addrspace(5) %"34", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 8
%"44" = load i16, ptr addrspace(5) %"36", align 2
%"46" = inttoptr i64 %"43" to ptr
store i16 %"44", ptr %"46", align 2

18
ptx/src/test/ll/cvt_rzi.ll
View File

@ -10,15 +10,15 @@ define amdgpu_kernel void @cvt_rzi(ptr addrspace(4) byref(i64) %"35", ptr addrsp
"34": ; preds = %1
call void @llvm.amdgcn.s.setreg(i32 6145, i32 3)
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(5) %"37", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(5) %"37", align 8
%"55" = inttoptr i64 %"44" to ptr
%"43" = load float, ptr %"55", align 4
store float %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"56" = inttoptr i64 %"45" to ptr
%"31" = getelementptr inbounds i8, ptr %"56", i64 4
%"46" = load float, ptr %"31", align 4
@ -31,11 +31,11 @@ define amdgpu_kernel void @cvt_rzi(ptr addrspace(4) byref(i64) %"35", ptr addrsp
%3 = call float @llvm.trunc.f32(float %"50")
%"49" = freeze float %3
store float %"49", ptr addrspace(5) %"40", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 8
%"52" = load float, ptr addrspace(5) %"39", align 4
%"57" = inttoptr i64 %"51" to ptr
store float %"52", ptr %"57", align 4
%"53" = load i64, ptr addrspace(5) %"38", align 4
%"53" = load i64, ptr addrspace(5) %"38", align 8
%"58" = inttoptr i64 %"53" to ptr
%"33" = getelementptr inbounds i8, ptr %"58", i64 4
%"54" = load float, ptr addrspace(5) %"40", align 4
@ -51,4 +51,4 @@ declare float @llvm.trunc.f32(float) #2
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="ieee" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #1 = { nocallback nofree nosync nounwind willreturn }
attributes #2 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
attributes #2 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }

12
ptx/src/test/ll/cvt_s16_s8.ll
View File

@ -9,11 +9,11 @@ define amdgpu_kernel void @cvt_s16_s8(ptr addrspace(4) byref(i64) %"31", ptr add
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"45" = inttoptr i64 %"40" to ptr addrspace(1)
%"39" = load i32, ptr addrspace(1) %"45", align 4
store i32 %"39", ptr addrspace(5) %"36", align 4
@ -22,7 +22,7 @@ define amdgpu_kernel void @cvt_s16_s8(ptr addrspace(4) byref(i64) %"31", ptr add
%"46" = sext i8 %2 to i16
%"41" = sext i16 %"46" to i32
store i32 %"41", ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 8
%"44" = load i32, ptr addrspace(5) %"35", align 4
%"48" = inttoptr i64 %"43" to ptr
store i32 %"44", ptr %"48", align 4

16
ptx/src/test/ll/cvt_s32_f32.ll
View File

@ -9,16 +9,16 @@ define amdgpu_kernel void @cvt_s32_f32(ptr addrspace(4) byref(i64) %"35", ptr ad
br label %"34"
"34": ; preds = %1
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(5) %"37", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(5) %"37", align 8
%"56" = inttoptr i64 %"44" to ptr
%"55" = load float, ptr %"56", align 4
%"43" = bitcast float %"55" to i32
store i32 %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"57" = inttoptr i64 %"45" to ptr
%"31" = getelementptr inbounds i8, ptr %"57", i64 4
%"58" = load float, ptr %"31", align 4
@ -36,11 +36,11 @@ define amdgpu_kernel void @cvt_s32_f32(ptr addrspace(4) byref(i64) %"35", ptr ad
%5 = fptosi float %4 to i32
%"61" = freeze i32 %5
store i32 %"61", ptr addrspace(5) %"40", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 8
%"52" = load i32, ptr addrspace(5) %"39", align 4
%"63" = inttoptr i64 %"51" to ptr addrspace(1)
store i32 %"52", ptr addrspace(1) %"63", align 4
%"53" = load i64, ptr addrspace(5) %"38", align 4
%"53" = load i64, ptr addrspace(5) %"38", align 8
%"65" = inttoptr i64 %"53" to ptr addrspace(1)
%"33" = getelementptr inbounds i8, ptr addrspace(1) %"65", i64 4
%"54" = load i32, ptr addrspace(5) %"40", align 4

18
ptx/src/test/ll/cvt_s64_s32.ll
View File

@ -9,21 +9,21 @@ define amdgpu_kernel void @cvt_s64_s32(ptr addrspace(4) byref(i64) %"31", ptr ad
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"46" = inttoptr i64 %"40" to ptr
%"45" = load i32, ptr %"46", align 4
store i32 %"45", ptr addrspace(5) %"35", align 4
%"42" = load i32, ptr addrspace(5) %"35", align 4
%"41" = sext i32 %"42" to i64
store i64 %"41", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 4
store i64 %"41", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"34", align 8
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"47" = inttoptr i64 %"43" to ptr
store i64 %"44", ptr %"47", align 4
store i64 %"44", ptr %"47", align 8
ret void
}

14
ptx/src/test/ll/cvt_sat_s_u.ll
View File

@ -10,11 +10,11 @@ define amdgpu_kernel void @cvt_sat_s_u(ptr addrspace(4) byref(i64) %"32", ptr ad
br label %"31"
"31": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"39", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"40", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"39", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"40", ptr addrspace(5) %"35", align 8
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"49" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"49", align 4
store i32 %"41", ptr addrspace(5) %"36", align 4
@ -24,7 +24,7 @@ define amdgpu_kernel void @cvt_sat_s_u(ptr addrspace(4) byref(i64) %"32", ptr ad
store i32 %3, ptr addrspace(5) %"37", align 4
%"46" = load i32, ptr addrspace(5) %"37", align 4
store i32 %"46", ptr addrspace(5) %"38", align 4
%"47" = load i64, ptr addrspace(5) %"35", align 4
%"47" = load i64, ptr addrspace(5) %"35", align 8
%"48" = load i32, ptr addrspace(5) %"38", align 4
%"50" = inttoptr i64 %"47" to ptr
store i32 %"48", ptr %"50", align 4
@ -38,4 +38,4 @@ declare i32 @llvm.smax.i32(i32, i32) #1
declare i32 @llvm.umin.i32(i32, i32) #1
attributes #0 = { "amdgpu-unsafe-fp-atomics"="true" "denormal-fp-math"="preserve-sign" "denormal-fp-math-f32"="preserve-sign" "no-trapping-math"="true" "uniform-work-group-size"="true" }
attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }

16
ptx/src/test/ll/cvta.ll
View File

@ -8,23 +8,23 @@ define amdgpu_kernel void @cvta(ptr addrspace(4) byref(i64) %"30", ptr addrspace
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%2 = inttoptr i64 %"38" to ptr
%"45" = addrspacecast ptr %2 to ptr addrspace(1)
store ptr addrspace(1) %"45", ptr addrspace(5) %"32", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 8
%3 = inttoptr i64 %"40" to ptr
%"47" = addrspacecast ptr %3 to ptr addrspace(1)
store ptr addrspace(1) %"47", ptr addrspace(5) %"33", align 8
%"42" = load i64, ptr addrspace(5) %"32", align 4
%"42" = load i64, ptr addrspace(5) %"32", align 8
%"49" = inttoptr i64 %"42" to ptr addrspace(1)
%"41" = load float, ptr addrspace(1) %"49", align 4
store float %"41", ptr addrspace(5) %"34", align 4
%"43" = load i64, ptr addrspace(5) %"33", align 4
%"43" = load i64, ptr addrspace(5) %"33", align 8
%"44" = load float, ptr addrspace(5) %"34", align 4
%"50" = inttoptr i64 %"43" to ptr addrspace(1)
store float %"44", ptr addrspace(1) %"50", align 4

14
ptx/src/test/ll/div_approx.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @div_approx(ptr addrspace(4) byref(i64) %"33", ptr add
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load float, ptr %"50", align 4
store float %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load float, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @div_approx(ptr addrspace(4) byref(i64) %"33", ptr add
%"47" = load float, ptr addrspace(5) %"38", align 4
%"45" = fdiv arcp afn float %"46", %"47"
store float %"45", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load float, ptr addrspace(5) %"37", align 4
%"52" = inttoptr i64 %"48" to ptr
store float %"49", ptr %"52", align 4

12
ptx/src/test/ll/ex2.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @ex2(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load float, ptr %"43", align 4
store float %"37", ptr addrspace(5) %"34", align 4
%"40" = load float, ptr addrspace(5) %"34", align 4
%"39" = call float @llvm.amdgcn.exp2.f32(float %"40")
store float %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load float, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store float %"42", ptr %"44", align 4

30
ptx/src/test/ll/extern_func.ll
View File

@ -13,29 +13,29 @@ define amdgpu_kernel void @extern_func(ptr addrspace(4) byref(i64) %"44", ptr ad
br label %"41"
"41": ; preds = %1
%"50" = load i64, ptr addrspace(4) %"44", align 4
store i64 %"50", ptr addrspace(5) %"46", align 4
%"51" = load i64, ptr addrspace(4) %"45", align 4
store i64 %"51", ptr addrspace(5) %"47", align 4
%"53" = load i64, ptr addrspace(5) %"46", align 4
%"50" = load i64, ptr addrspace(4) %"44", align 8
store i64 %"50", ptr addrspace(5) %"46", align 8
%"51" = load i64, ptr addrspace(4) %"45", align 8
store i64 %"51", ptr addrspace(5) %"47", align 8
%"53" = load i64, ptr addrspace(5) %"46", align 8
%"61" = inttoptr i64 %"53" to ptr addrspace(1)
%"52" = load i64, ptr addrspace(1) %"61", align 4
store i64 %"52", ptr addrspace(5) %"48", align 4
%"52" = load i64, ptr addrspace(1) %"61", align 8
store i64 %"52", ptr addrspace(5) %"48", align 8
%"55" = getelementptr inbounds i8, ptr addrspace(5) %"54", i64 0
%"56" = load i64, ptr addrspace(5) %"48", align 4
store i64 %"56", ptr addrspace(5) %"55", align 4
%"39" = load i64, ptr addrspace(5) %"54", align 4
%"56" = load i64, ptr addrspace(5) %"48", align 8
store i64 %"56", ptr addrspace(5) %"55", align 8
%"39" = load i64, ptr addrspace(5) %"54", align 8
%"40" = call [16 x i8] @foobar(i64 %"39")
br label %"42"
"42": ; preds = %"41"
store [16 x i8] %"40", ptr addrspace(5) %"57", align 1
%"58" = load i64, ptr addrspace(5) %"57", align 4
store i64 %"58", ptr addrspace(5) %"49", align 4
%"59" = load i64, ptr addrspace(5) %"47", align 4
%"60" = load i64, ptr addrspace(5) %"49", align 4
%"58" = load i64, ptr addrspace(5) %"57", align 8
store i64 %"58", ptr addrspace(5) %"49", align 8
%"59" = load i64, ptr addrspace(5) %"47", align 8
%"60" = load i64, ptr addrspace(5) %"49", align 8
%"64" = inttoptr i64 %"59" to ptr
store i64 %"60", ptr %"64", align 4
store i64 %"60", ptr %"64", align 8
ret void
}

28
ptx/src/test/ll/extern_shared.ll
View File

@ -10,22 +10,22 @@ define amdgpu_kernel void @extern_shared(ptr addrspace(4) byref(i64) %"31", ptr
br label %"30"
"30": ; preds = %1
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"44" = inttoptr i64 %"39" to ptr addrspace(1)
%"38" = load i64, ptr addrspace(1) %"44", align 4
store i64 %"38", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(5) %"35", align 4
store i64 %"40", ptr addrspace(3) @shared_mem, align 4
%"41" = load i64, ptr addrspace(3) @shared_mem, align 4
store i64 %"41", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"38" = load i64, ptr addrspace(1) %"44", align 8
store i64 %"38", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(5) %"35", align 8
store i64 %"40", ptr addrspace(3) @shared_mem, align 8
%"41" = load i64, ptr addrspace(3) @shared_mem, align 8
store i64 %"41", ptr addrspace(5) %"35", align 8
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"47" = inttoptr i64 %"42" to ptr addrspace(1)
store i64 %"43", ptr addrspace(1) %"47", align 4
store i64 %"43", ptr addrspace(1) %"47", align 8
ret void
}

40
ptx/src/test/ll/extern_shared_call.ll
View File

@ -8,13 +8,13 @@ define void @incr_shared_2_global() #0 {
br label %"33"
"33": ; preds = %1
%"37" = load i64, ptr addrspace(3) @shared_mem, align 4
store i64 %"37", ptr addrspace(5) %"36", align 4
%"39" = load i64, ptr addrspace(5) %"36", align 4
%"37" = load i64, ptr addrspace(3) @shared_mem, align 8
store i64 %"37", ptr addrspace(5) %"36", align 8
%"39" = load i64, ptr addrspace(5) %"36", align 8
%"38" = add i64 %"39", 2
store i64 %"38", ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr addrspace(5) %"36", align 4
store i64 %"40", ptr addrspace(3) @shared_mem, align 4
store i64 %"38", ptr addrspace(5) %"36", align 8
%"40" = load i64, ptr addrspace(5) %"36", align 8
store i64 %"40", ptr addrspace(3) @shared_mem, align 8
ret void
}
@ -28,26 +28,26 @@ define amdgpu_kernel void @extern_shared_call(ptr addrspace(4) byref(i64) %"41",
br label %"34"
"34": ; preds = %1
%"46" = load i64, ptr addrspace(4) %"41", align 4
store i64 %"46", ptr addrspace(5) %"43", align 4
%"47" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"47", ptr addrspace(5) %"44", align 4
%"49" = load i64, ptr addrspace(5) %"43", align 4
%"46" = load i64, ptr addrspace(4) %"41", align 8
store i64 %"46", ptr addrspace(5) %"43", align 8
%"47" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"47", ptr addrspace(5) %"44", align 8
%"49" = load i64, ptr addrspace(5) %"43", align 8
%"56" = inttoptr i64 %"49" to ptr addrspace(1)
%"48" = load i64, ptr addrspace(1) %"56", align 4
store i64 %"48", ptr addrspace(5) %"45", align 4
%"50" = load i64, ptr addrspace(5) %"45", align 4
store i64 %"50", ptr addrspace(3) @shared_mem, align 4
%"48" = load i64, ptr addrspace(1) %"56", align 8
store i64 %"48", ptr addrspace(5) %"45", align 8
%"50" = load i64, ptr addrspace(5) %"45", align 8
store i64 %"50", ptr addrspace(3) @shared_mem, align 8
call void @incr_shared_2_global()
br label %"35"
"35": ; preds = %"34"
%"51" = load i64, ptr addrspace(3) @shared_mem, align 4
store i64 %"51", ptr addrspace(5) %"45", align 4
%"52" = load i64, ptr addrspace(5) %"44", align 4
%"53" = load i64, ptr addrspace(5) %"45", align 4
%"51" = load i64, ptr addrspace(3) @shared_mem, align 8
store i64 %"51", ptr addrspace(5) %"45", align 8
%"52" = load i64, ptr addrspace(5) %"44", align 8
%"53" = load i64, ptr addrspace(5) %"45", align 8
%"59" = inttoptr i64 %"52" to ptr addrspace(1)
store i64 %"53", ptr addrspace(1) %"59", align 4
store i64 %"53", ptr addrspace(1) %"59", align 8
ret void
}

16
ptx/src/test/ll/fma.ll
View File

@ -10,20 +10,20 @@ define amdgpu_kernel void @fma(ptr addrspace(4) byref(i64) %"36", ptr addrspace(
br label %"35"
"35": ; preds = %1
%"43" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"43", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"44", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"43", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"44", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(5) %"38", align 8
%"57" = inttoptr i64 %"46" to ptr
%"45" = load float, ptr %"57", align 4
store float %"45", ptr addrspace(5) %"40", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 8
%"58" = inttoptr i64 %"47" to ptr
%"32" = getelementptr inbounds i8, ptr %"58", i64 4
%"48" = load float, ptr %"32", align 4
store float %"48", ptr addrspace(5) %"41", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 8
%"59" = inttoptr i64 %"49" to ptr
%"34" = getelementptr inbounds i8, ptr %"59", i64 8
%"50" = load float, ptr %"34", align 4
@ -33,7 +33,7 @@ define amdgpu_kernel void @fma(ptr addrspace(4) byref(i64) %"36", ptr addrspace(
%"54" = load float, ptr addrspace(5) %"42", align 4
%"51" = call float @llvm.fma.f32(float %"52", float %"53", float %"54")
store float %"51", ptr addrspace(5) %"40", align 4
%"55" = load i64, ptr addrspace(5) %"39", align 4
%"55" = load i64, ptr addrspace(5) %"39", align 8
%"56" = load float, ptr addrspace(5) %"40", align 4
%"60" = inttoptr i64 %"55" to ptr
store float %"56", ptr %"60", align 4

14
ptx/src/test/ll/fmax.ll
View File

@ -11,16 +11,16 @@ define amdgpu_kernel void @fmax(ptr addrspace(4) byref(i64) %"35", ptr addrspace
br label %"34"
"34": ; preds = %1
%"43" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"43", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"44", ptr addrspace(5) %"38", align 4
%"46" = load i64, ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"43", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"44", ptr addrspace(5) %"38", align 8
%"46" = load i64, ptr addrspace(5) %"37", align 8
%"55" = inttoptr i64 %"46" to ptr
%"54" = load i16, ptr %"55", align 2
%"45" = bitcast i16 %"54" to half
store half %"45", ptr addrspace(5) %"39", align 2
%"47" = load i64, ptr addrspace(5) %"37", align 4
%"47" = load i64, ptr addrspace(5) %"37", align 8
%"56" = inttoptr i64 %"47" to ptr
%"33" = getelementptr inbounds i8, ptr %"56", i64 2
%"57" = load i16, ptr %"33", align 2
@ -30,7 +30,7 @@ define amdgpu_kernel void @fmax(ptr addrspace(4) byref(i64) %"35", ptr addrspace
%"51" = load half, ptr addrspace(5) %"39", align 2
%"49" = call half @llvm.maxnum.f16(half %"50", half %"51")
store half %"49", ptr addrspace(5) %"41", align 2
%"52" = load i64, ptr addrspace(5) %"38", align 4
%"52" = load i64, ptr addrspace(5) %"38", align 8
%"53" = load half, ptr addrspace(5) %"41", align 2
%"58" = inttoptr i64 %"52" to ptr
%"59" = bitcast half %"53" to i16

10
ptx/src/test/ll/global_array.ll
View File

@ -10,14 +10,14 @@ define amdgpu_kernel void @global_array(ptr addrspace(4) byref(i64) %"31", ptr a
br label %"30"
"30": ; preds = %1
store i64 ptrtoint (ptr addrspace(1) @foobar to i64), ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
store i64 ptrtoint (ptr addrspace(1) @foobar to i64), ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"43" = inttoptr i64 %"39" to ptr addrspace(1)
%"38" = load i32, ptr addrspace(1) %"43", align 4
store i32 %"38", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"34", align 8
%"41" = load i32, ptr addrspace(5) %"35", align 4
%"44" = inttoptr i64 %"40" to ptr addrspace(1)
store i32 %"41", ptr addrspace(1) %"44", align 4

12
ptx/src/test/ll/lanemask_lt.ll
View File

@ -12,11 +12,11 @@ define amdgpu_kernel void @lanemask_lt(ptr addrspace(4) byref(i64) %"36", ptr ad
br label %"33"
"33": ; preds = %1
%"43" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"43", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"44", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"43", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"44", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(5) %"38", align 8
%"56" = inttoptr i64 %"46" to ptr
%"55" = load i32, ptr %"56", align 4
store i32 %"55", ptr addrspace(5) %"40", align 4
@ -32,7 +32,7 @@ define amdgpu_kernel void @lanemask_lt(ptr addrspace(4) byref(i64) %"36", ptr ad
%"52" = load i32, ptr addrspace(5) %"42", align 4
%"60" = add i32 %"51", %"52"
store i32 %"60", ptr addrspace(5) %"41", align 4
%"53" = load i64, ptr addrspace(5) %"39", align 4
%"53" = load i64, ptr addrspace(5) %"39", align 8
%"54" = load i32, ptr addrspace(5) %"41", align 4
%"63" = inttoptr i64 %"53" to ptr
store i32 %"54", ptr %"63", align 4

20
ptx/src/test/ll/ld_st.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @ld_st(ptr addrspace(4) byref(i64) %"30", ptr addrspac
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"41" = inttoptr i64 %"38" to ptr
%"37" = load i64, ptr %"41", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"40" = load i64, ptr addrspace(5) %"34", align 4
%"37" = load i64, ptr %"41", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"40" = load i64, ptr addrspace(5) %"34", align 8
%"42" = inttoptr i64 %"39" to ptr
store i64 %"40", ptr %"42", align 4
store i64 %"40", ptr %"42", align 8
ret void
}

18
ptx/src/test/ll/ld_st_implicit.ll
View File

@ -8,19 +8,19 @@ define amdgpu_kernel void @ld_st_implicit(ptr addrspace(4) byref(i64) %"31", ptr
br label %"30"
"30": ; preds = %1
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
store i64 81985529216486895, ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
store i64 81985529216486895, ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"44" = inttoptr i64 %"40" to ptr addrspace(1)
%"43" = load float, ptr addrspace(1) %"44", align 4
%2 = bitcast float %"43" to i32
%"39" = zext i32 %2 to i64
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"45" = inttoptr i64 %"41" to ptr addrspace(1)
%3 = trunc i64 %"42" to i32
%"46" = bitcast i32 %3 to float

16
ptx/src/test/ll/ld_st_offset.ll
View File

@ -9,24 +9,24 @@ define amdgpu_kernel void @ld_st_offset(ptr addrspace(4) byref(i64) %"35", ptr a
br label %"34"
"34": ; preds = %1
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(5) %"37", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(5) %"37", align 8
%"51" = inttoptr i64 %"44" to ptr
%"43" = load i32, ptr %"51", align 4
store i32 %"43", ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"52" = inttoptr i64 %"45" to ptr
%"31" = getelementptr inbounds i8, ptr %"52", i64 4
%"46" = load i32, ptr %"31", align 4
store i32 %"46", ptr addrspace(5) %"40", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 8
%"48" = load i32, ptr addrspace(5) %"40", align 4
%"53" = inttoptr i64 %"47" to ptr
store i32 %"48", ptr %"53", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 8
%"54" = inttoptr i64 %"49" to ptr
%"33" = getelementptr inbounds i8, ptr %"54", i64 4
%"50" = load i32, ptr addrspace(5) %"39", align 4

12
ptx/src/test/ll/lg2.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @lg2(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load float, ptr %"43", align 4
store float %"37", ptr addrspace(5) %"34", align 4
%"40" = load float, ptr addrspace(5) %"34", align 4
%"39" = call float @llvm.amdgcn.log.f32(float %"40")
store float %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load float, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store float %"42", ptr %"44", align 4

20
ptx/src/test/ll/local_align.ll
View File

@ -9,18 +9,18 @@ define amdgpu_kernel void @local_align(ptr addrspace(4) byref(i64) %"31", ptr ad
br label %"30"
"30": ; preds = %1
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"42" = inttoptr i64 %"39" to ptr
%"38" = load i64, ptr %"42", align 4
store i64 %"38", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"35", align 4
%"38" = load i64, ptr %"42", align 8
store i64 %"38", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(5) %"34", align 8
%"41" = load i64, ptr addrspace(5) %"35", align 8
%"43" = inttoptr i64 %"40" to ptr
store i64 %"41", ptr %"43", align 4
store i64 %"41", ptr %"43", align 8
ret void
}

16
ptx/src/test/ll/mad_s32.ll
View File

@ -11,20 +11,20 @@ define amdgpu_kernel void @mad_s32(ptr addrspace(4) byref(i64) %"37", ptr addrsp
br label %"36"
"36": ; preds = %1
%"45" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"45", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(4) %"38", align 4
store i64 %"46", ptr addrspace(5) %"40", align 4
%"48" = load i64, ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"45", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(4) %"38", align 8
store i64 %"46", ptr addrspace(5) %"40", align 8
%"48" = load i64, ptr addrspace(5) %"39", align 8
%"59" = inttoptr i64 %"48" to ptr
%"47" = load i32, ptr %"59", align 4
store i32 %"47", ptr addrspace(5) %"42", align 4
%"49" = load i64, ptr addrspace(5) %"39", align 4
%"49" = load i64, ptr addrspace(5) %"39", align 8
%"60" = inttoptr i64 %"49" to ptr
%"33" = getelementptr inbounds i8, ptr %"60", i64 4
%"50" = load i32, ptr %"33", align 4
store i32 %"50", ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr addrspace(5) %"39", align 4
%"51" = load i64, ptr addrspace(5) %"39", align 8
%"61" = inttoptr i64 %"51" to ptr
%"35" = getelementptr inbounds i8, ptr %"61", i64 8
%"52" = load i32, ptr %"35", align 4
@ -35,7 +35,7 @@ define amdgpu_kernel void @mad_s32(ptr addrspace(4) byref(i64) %"37", ptr addrsp
%2 = mul i32 %"54", %"55"
%"53" = add i32 %2, %"56"
store i32 %"53", ptr addrspace(5) %"41", align 4
%"57" = load i64, ptr addrspace(5) %"40", align 4
%"57" = load i64, ptr addrspace(5) %"40", align 8
%"58" = load i32, ptr addrspace(5) %"41", align 4
%"62" = inttoptr i64 %"57" to ptr
store i32 %"58", ptr %"62", align 4

28
ptx/src/test/ll/mad_wide.ll
View File

@ -11,36 +11,36 @@ define amdgpu_kernel void @mad_wide(ptr addrspace(4) byref(i64) %"37", ptr addrs
br label %"36"
"36": ; preds = %1
%"45" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"45", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(4) %"38", align 4
store i64 %"46", ptr addrspace(5) %"40", align 4
%"48" = load i64, ptr addrspace(5) %"39", align 4
%"45" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"45", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(4) %"38", align 8
store i64 %"46", ptr addrspace(5) %"40", align 8
%"48" = load i64, ptr addrspace(5) %"39", align 8
%"59" = inttoptr i64 %"48" to ptr
%"47" = load i32, ptr %"59", align 4
store i32 %"47", ptr addrspace(5) %"42", align 4
%"49" = load i64, ptr addrspace(5) %"39", align 4
%"49" = load i64, ptr addrspace(5) %"39", align 8
%"60" = inttoptr i64 %"49" to ptr
%"33" = getelementptr inbounds i8, ptr %"60", i64 4
%"50" = load i32, ptr %"33", align 4
store i32 %"50", ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr addrspace(5) %"39", align 4
%"51" = load i64, ptr addrspace(5) %"39", align 8
%"61" = inttoptr i64 %"51" to ptr
%"35" = getelementptr inbounds i8, ptr %"61", i64 8
%"52" = load i64, ptr %"35", align 4
store i64 %"52", ptr addrspace(5) %"44", align 4
%"52" = load i64, ptr %"35", align 8
store i64 %"52", ptr addrspace(5) %"44", align 8
%"54" = load i32, ptr addrspace(5) %"42", align 4
%"55" = load i32, ptr addrspace(5) %"43", align 4
%"56" = load i64, ptr addrspace(5) %"44", align 4
%"56" = load i64, ptr addrspace(5) %"44", align 8
%2 = sext i32 %"54" to i64
%3 = sext i32 %"55" to i64
%4 = mul i64 %2, %3
%"53" = add i64 %4, %"56"
store i64 %"53", ptr addrspace(5) %"41", align 4
%"57" = load i64, ptr addrspace(5) %"40", align 4
%"58" = load i64, ptr addrspace(5) %"41", align 4
store i64 %"53", ptr addrspace(5) %"41", align 8
%"57" = load i64, ptr addrspace(5) %"40", align 8
%"58" = load i64, ptr addrspace(5) %"41", align 8
%"62" = inttoptr i64 %"57" to ptr
store i64 %"58", ptr %"62", align 4
store i64 %"58", ptr %"62", align 8
ret void
}

24
ptx/src/test/ll/malformed_label.ll
View File

@ -9,24 +9,24 @@ define amdgpu_kernel void @malformed_label(ptr addrspace(4) byref(i64) %"34", pt
br label %"32"
"32": ; preds = %1
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
br label %"10"
"10": ; preds = %"32"
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"48" = inttoptr i64 %"43" to ptr
%"42" = load i64, ptr %"48", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"45" = load i64, ptr addrspace(5) %"38", align 4
%"42" = load i64, ptr %"48", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"45" = load i64, ptr addrspace(5) %"38", align 8
%"44" = add i64 %"45", 1
store i64 %"44", ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(5) %"37", align 4
%"47" = load i64, ptr addrspace(5) %"39", align 4
store i64 %"44", ptr addrspace(5) %"39", align 8
%"46" = load i64, ptr addrspace(5) %"37", align 8
%"47" = load i64, ptr addrspace(5) %"39", align 8
%"49" = inttoptr i64 %"46" to ptr
store i64 %"47", ptr %"49", align 4
store i64 %"47", ptr %"49", align 8
ret void
}

14
ptx/src/test/ll/max.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @max(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load i32, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @max(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
%"47" = load i32, ptr addrspace(5) %"38", align 4
%"45" = call i32 @llvm.smax.i32(i32 %"46", i32 %"47")
store i32 %"45", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load i32, ptr addrspace(5) %"37", align 4
%"52" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"52", align 4

12
ptx/src/test/ll/membar.ll
View File

@ -8,16 +8,16 @@ define amdgpu_kernel void @membar(ptr addrspace(4) byref(i64) %"30", ptr addrspa
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"42" = inttoptr i64 %"38" to ptr
%"41" = load i32, ptr %"42", align 4
store i32 %"41", ptr addrspace(5) %"34", align 4
fence seq_cst
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"40" = load i32, ptr addrspace(5) %"34", align 4
%"43" = inttoptr i64 %"39" to ptr
store i32 %"40", ptr %"43", align 4

14
ptx/src/test/ll/min.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @min(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load i32, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @min(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
%"47" = load i32, ptr addrspace(5) %"38", align 4
%"45" = call i32 @llvm.smin.i32(i32 %"46", i32 %"47")
store i32 %"45", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load i32, ptr addrspace(5) %"37", align 4
%"52" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"52", align 4

24
ptx/src/test/ll/mov.ll
View File

@ -9,20 +9,20 @@ define amdgpu_kernel void @mov(ptr addrspace(4) byref(i64) %"31", ptr addrspace(
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"45" = inttoptr i64 %"40" to ptr
%"39" = load i64, ptr %"45", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
store i64 %"42", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 4
%"39" = load i64, ptr %"45", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
store i64 %"42", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"34", align 8
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"46" = inttoptr i64 %"43" to ptr
store i64 %"44", ptr %"46", align 4
store i64 %"44", ptr %"46", align 8
ret void
}

2
ptx/src/test/ll/mov_address.ll
View File

@ -8,7 +8,7 @@ define amdgpu_kernel void @mov_address(ptr addrspace(4) byref(i64) %"29", ptr ad
"28": ; preds = %1
%"33" = ptrtoint ptr addrspace(5) %"10" to i64
store i64 %"33", ptr addrspace(5) %"31", align 4
store i64 %"33", ptr addrspace(5) %"31", align 8
ret void
}

12
ptx/src/test/ll/mul24_hi_s32.ll
View File

@ -10,11 +10,11 @@ define amdgpu_kernel void @mul24_hi_s32(ptr addrspace(4) byref(i64) %"32", ptr a
br label %"31"
"31": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"39", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"40", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"39", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"40", ptr addrspace(5) %"35", align 8
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"36", align 4
@ -29,7 +29,7 @@ define amdgpu_kernel void @mul24_hi_s32(ptr addrspace(4) byref(i64) %"32", ptr a
%5 = shl i32 %3, 16
%"45" = or i32 %4, %5
store i32 %"45", ptr addrspace(5) %"38", align 4
%"48" = load i64, ptr addrspace(5) %"35", align 4
%"48" = load i64, ptr addrspace(5) %"35", align 8
%"49" = load i32, ptr addrspace(5) %"38", align 4
%"51" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"51", align 4

12
ptx/src/test/ll/mul24_hi_u32.ll
View File

@ -9,11 +9,11 @@ define amdgpu_kernel void @mul24_hi_u32(ptr addrspace(4) byref(i64) %"31", ptr a
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"46" = inttoptr i64 %"40" to ptr
%"39" = load i32, ptr %"46", align 4
store i32 %"39", ptr addrspace(5) %"35", align 4
@ -25,7 +25,7 @@ define amdgpu_kernel void @mul24_hi_u32(ptr addrspace(4) byref(i64) %"31", ptr a
%5 = shl i32 %3, 16
%"41" = or i32 %4, %5
store i32 %"41", ptr addrspace(5) %"36", align 4
%"44" = load i64, ptr addrspace(5) %"34", align 4
%"44" = load i64, ptr addrspace(5) %"34", align 8
%"45" = load i32, ptr addrspace(5) %"36", align 4
%"47" = inttoptr i64 %"44" to ptr
store i32 %"45", ptr %"47", align 4

12
ptx/src/test/ll/mul24_lo_s32.ll
View File

@ -10,11 +10,11 @@ define amdgpu_kernel void @mul24_lo_s32(ptr addrspace(4) byref(i64) %"32", ptr a
br label %"31"
"31": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"39", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"40", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"39", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"40", ptr addrspace(5) %"35", align 8
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"36", align 4
@ -25,7 +25,7 @@ define amdgpu_kernel void @mul24_lo_s32(ptr addrspace(4) byref(i64) %"32", ptr a
%"47" = load i32, ptr addrspace(5) %"36", align 4
%"45" = call i32 @llvm.amdgcn.mul.i24(i32 %"46", i32 %"47")
store i32 %"45", ptr addrspace(5) %"38", align 4
%"48" = load i64, ptr addrspace(5) %"35", align 4
%"48" = load i64, ptr addrspace(5) %"35", align 8
%"49" = load i32, ptr addrspace(5) %"38", align 4
%"51" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"51", align 4

12
ptx/src/test/ll/mul24_lo_u32.ll
View File

@ -9,11 +9,11 @@ define amdgpu_kernel void @mul24_lo_u32(ptr addrspace(4) byref(i64) %"31", ptr a
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"46" = inttoptr i64 %"40" to ptr
%"39" = load i32, ptr %"46", align 4
store i32 %"39", ptr addrspace(5) %"35", align 4
@ -21,7 +21,7 @@ define amdgpu_kernel void @mul24_lo_u32(ptr addrspace(4) byref(i64) %"31", ptr a
%"43" = load i32, ptr addrspace(5) %"35", align 4
%"41" = call i32 @llvm.amdgcn.mul.u24(i32 %"42", i32 %"43")
store i32 %"41", ptr addrspace(5) %"36", align 4
%"44" = load i64, ptr addrspace(5) %"34", align 4
%"44" = load i64, ptr addrspace(5) %"34", align 8
%"45" = load i32, ptr addrspace(5) %"36", align 4
%"47" = inttoptr i64 %"44" to ptr
store i32 %"45", ptr %"47", align 4

14
ptx/src/test/ll/mul_ftz.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @mul_ftz(ptr addrspace(4) byref(i64) %"33", ptr addrsp
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load float, ptr %"50", align 4
store float %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load float, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @mul_ftz(ptr addrspace(4) byref(i64) %"33", ptr addrsp
%"47" = load float, ptr addrspace(5) %"38", align 4
%"45" = fmul float %"46", %"47"
store float %"45", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load float, ptr addrspace(5) %"37", align 4
%"52" = inttoptr i64 %"48" to ptr
store float %"49", ptr %"52", align 4

24
ptx/src/test/ll/mul_hi.ll
View File

@ -9,24 +9,24 @@ define amdgpu_kernel void @mul_hi(ptr addrspace(4) byref(i64) %"32", ptr addrspa
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr
%"40" = load i64, ptr %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%2 = zext i64 %"43" to i128
%3 = mul i128 %2, 2
%4 = lshr i128 %3, 64
%"42" = trunc i128 %4 to i64
store i64 %"42", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"42", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"47" = inttoptr i64 %"44" to ptr
store i64 %"45", ptr %"47", align 4
store i64 %"45", ptr %"47", align 8
ret void
}

24
ptx/src/test/ll/mul_lo.ll
View File

@ -9,21 +9,21 @@ define amdgpu_kernel void @mul_lo(ptr addrspace(4) byref(i64) %"32", ptr addrspa
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr
%"40" = load i64, ptr %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"42" = mul i64 %"43", 2
store i64 %"42", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"42", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"47" = inttoptr i64 %"44" to ptr
store i64 %"45", ptr %"47", align 4
store i64 %"45", ptr %"47", align 8
ret void
}

14
ptx/src/test/ll/mul_non_ftz.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @mul_non_ftz(ptr addrspace(4) byref(i64) %"33", ptr ad
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load float, ptr %"50", align 4
store float %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load float, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @mul_non_ftz(ptr addrspace(4) byref(i64) %"33", ptr ad
%"47" = load float, ptr addrspace(5) %"38", align 4
%"45" = fmul float %"46", %"47"
store float %"45", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load float, ptr addrspace(5) %"37", align 4
%"52" = inttoptr i64 %"48" to ptr
store float %"49", ptr %"52", align 4

20
ptx/src/test/ll/mul_wide.ll
View File

@ -10,15 +10,15 @@ define amdgpu_kernel void @mul_wide(ptr addrspace(4) byref(i64) %"34", ptr addrs
br label %"33"
"33": ; preds = %1
%"41" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"41", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"42", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"41", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"42", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"52" = inttoptr i64 %"44" to ptr addrspace(1)
%"43" = load i32, ptr addrspace(1) %"52", align 4
store i32 %"43", ptr addrspace(5) %"38", align 4
%"45" = load i64, ptr addrspace(5) %"36", align 4
%"45" = load i64, ptr addrspace(5) %"36", align 8
%"53" = inttoptr i64 %"45" to ptr addrspace(1)
%"32" = getelementptr inbounds i8, ptr addrspace(1) %"53", i64 4
%"46" = load i32, ptr addrspace(1) %"32", align 4
@ -28,11 +28,11 @@ define amdgpu_kernel void @mul_wide(ptr addrspace(4) byref(i64) %"34", ptr addrs
%2 = sext i32 %"48" to i64
%3 = sext i32 %"49" to i64
%"47" = mul i64 %2, %3
store i64 %"47", ptr addrspace(5) %"40", align 4
%"50" = load i64, ptr addrspace(5) %"37", align 4
%"51" = load i64, ptr addrspace(5) %"40", align 4
store i64 %"47", ptr addrspace(5) %"40", align 8
%"50" = load i64, ptr addrspace(5) %"37", align 8
%"51" = load i64, ptr addrspace(5) %"40", align 8
%"54" = inttoptr i64 %"50" to ptr
store i64 %"51", ptr %"54", align 4
store i64 %"51", ptr %"54", align 8
ret void
}

14
ptx/src/test/ll/multiple_return.ll
View File

@ -30,11 +30,11 @@ define amdgpu_kernel void @multiple_return(ptr addrspace(4) byref(i64) %"50", pt
br label %"44"
"44": ; preds = %1
%"57" = load i64, ptr addrspace(4) %"50", align 4
store i64 %"57", ptr addrspace(5) %"52", align 4
%"58" = load i64, ptr addrspace(4) %"51", align 4
store i64 %"58", ptr addrspace(5) %"53", align 4
%"60" = load i64, ptr addrspace(5) %"52", align 4
%"57" = load i64, ptr addrspace(4) %"50", align 8
store i64 %"57", ptr addrspace(5) %"52", align 8
%"58" = load i64, ptr addrspace(4) %"51", align 8
store i64 %"58", ptr addrspace(5) %"53", align 8
%"60" = load i64, ptr addrspace(5) %"52", align 8
%"68" = inttoptr i64 %"60" to ptr
%"59" = load i32, ptr %"68", align 4
store i32 %"59", ptr addrspace(5) %"54", align 4
@ -48,7 +48,7 @@ define amdgpu_kernel void @multiple_return(ptr addrspace(4) byref(i64) %"50", pt
br label %"45"
"45": ; preds = %"44"
%"64" = load i64, ptr addrspace(5) %"53", align 4
%"64" = load i64, ptr addrspace(5) %"53", align 8
%"65" = load i32, ptr addrspace(5) %"55", align 4
%"69" = inttoptr i64 %"64" to ptr
store i32 %"65", ptr %"69", align 4
@ -56,7 +56,7 @@ define amdgpu_kernel void @multiple_return(ptr addrspace(4) byref(i64) %"50", pt
br i1 %"66", label %"19", label %"20"
"19": ; preds = %"45"
%"67" = load i64, ptr addrspace(5) %"53", align 4
%"67" = load i64, ptr addrspace(5) %"53", align 8
%"70" = inttoptr i64 %"67" to ptr
%"41" = getelementptr inbounds i8, ptr %"70", i64 4
store i32 123, ptr %"41", align 4

12
ptx/src/test/ll/neg.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @neg(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load i32, ptr %"43", align 4
store i32 %"37", ptr addrspace(5) %"34", align 4
%"40" = load i32, ptr addrspace(5) %"34", align 4
%"39" = sub i32 0, %"40"
store i32 %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load i32, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store i32 %"42", ptr %"44", align 4

12
ptx/src/test/ll/non_scalar_ptr_offset.ll
View File

@ -9,11 +9,11 @@ define amdgpu_kernel void @non_scalar_ptr_offset(ptr addrspace(4) byref(i64) %"3
br label %"33"
"33": ; preds = %1
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"42" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"42" = load i64, ptr addrspace(5) %"36", align 8
%"50" = inttoptr i64 %"42" to ptr addrspace(1)
%"32" = getelementptr inbounds i8, ptr addrspace(1) %"50", i64 8
%"30" = load <2 x i32>, ptr addrspace(1) %"32", align 8
@ -25,7 +25,7 @@ define amdgpu_kernel void @non_scalar_ptr_offset(ptr addrspace(4) byref(i64) %"3
%"47" = load i32, ptr addrspace(5) %"39", align 4
%"45" = add i32 %"46", %"47"
store i32 %"45", ptr addrspace(5) %"38", align 4
%"48" = load i64, ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"37", align 8
%"49" = load i32, ptr addrspace(5) %"38", align 4
%"51" = inttoptr i64 %"48" to ptr addrspace(1)
store i32 %"49", ptr addrspace(1) %"51", align 4

24
ptx/src/test/ll/not.ll
View File

@ -9,21 +9,21 @@ define amdgpu_kernel void @not(ptr addrspace(4) byref(i64) %"31", ptr addrspace(
br label %"30"
"30": ; preds = %1
%"37" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"37", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"40" = load i64, ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"37", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"40" = load i64, ptr addrspace(5) %"33", align 8
%"45" = inttoptr i64 %"40" to ptr
%"39" = load i64, ptr %"45", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr %"45", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"46" = xor i64 %"42", -1
store i64 %"46", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"34", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 4
store i64 %"46", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"34", align 8
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"48" = inttoptr i64 %"43" to ptr
store i64 %"44", ptr %"48", align 4
store i64 %"44", ptr %"48", align 8
ret void
}

12
ptx/src/test/ll/ntid.ll
View File

@ -11,11 +11,11 @@ define amdgpu_kernel void @ntid(ptr addrspace(4) byref(i64) %"35", ptr addrspace
br label %"32"
"32": ; preds = %1
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"42" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"42", ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr addrspace(5) %"37", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"42" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"42", ptr addrspace(5) %"38", align 8
%"44" = load i64, ptr addrspace(5) %"37", align 8
%"51" = inttoptr i64 %"44" to ptr
%"43" = load i32, ptr %"51", align 4
store i32 %"43", ptr addrspace(5) %"39", align 4
@ -28,7 +28,7 @@ define amdgpu_kernel void @ntid(ptr addrspace(4) byref(i64) %"35", ptr addrspace
%"48" = load i32, ptr addrspace(5) %"40", align 4
%"46" = add i32 %"47", %"48"
store i32 %"46", ptr addrspace(5) %"39", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 4
%"49" = load i64, ptr addrspace(5) %"38", align 8
%"50" = load i32, ptr addrspace(5) %"39", align 4
%"52" = inttoptr i64 %"49" to ptr
store i32 %"50", ptr %"52", align 4

32
ptx/src/test/ll/or.ll
View File

@ -9,27 +9,27 @@ define amdgpu_kernel void @or(ptr addrspace(4) byref(i64) %"33", ptr addrspace(4
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i64, ptr %"50", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr %"50", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 8
%"44" = load i64, ptr %"31", align 4
store i64 %"44", ptr addrspace(5) %"38", align 4
%"46" = load i64, ptr addrspace(5) %"37", align 4
%"47" = load i64, ptr addrspace(5) %"38", align 4
%"44" = load i64, ptr %"31", align 8
store i64 %"44", ptr addrspace(5) %"38", align 8
%"46" = load i64, ptr addrspace(5) %"37", align 8
%"47" = load i64, ptr addrspace(5) %"38", align 8
%"52" = or i64 %"46", %"47"
store i64 %"52", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"49" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"52", ptr addrspace(5) %"37", align 8
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load i64, ptr addrspace(5) %"37", align 8
%"55" = inttoptr i64 %"48" to ptr
store i64 %"49", ptr %"55", align 4
store i64 %"49", ptr %"55", align 8
ret void
}

12
ptx/src/test/ll/popc.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @popc(ptr addrspace(4) byref(i64) %"30", ptr addrspace
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load i32, ptr %"43", align 4
store i32 %"37", ptr addrspace(5) %"34", align 4
%"40" = load i32, ptr addrspace(5) %"34", align 4
%"44" = call i32 @llvm.ctpop.i32(i32 %"40")
store i32 %"44", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load i32, ptr addrspace(5) %"34", align 4
%"45" = inttoptr i64 %"41" to ptr
store i32 %"42", ptr %"45", align 4

34
ptx/src/test/ll/pred_not.ll
View File

@ -11,21 +11,21 @@ define amdgpu_kernel void @pred_not(ptr addrspace(4) byref(i64) %"41", ptr addrs
br label %"40"
"40": ; preds = %1
%"49" = load i64, ptr addrspace(4) %"41", align 4
store i64 %"49", ptr addrspace(5) %"43", align 4
%"50" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"50", ptr addrspace(5) %"44", align 4
%"52" = load i64, ptr addrspace(5) %"43", align 4
%"49" = load i64, ptr addrspace(4) %"41", align 8
store i64 %"49", ptr addrspace(5) %"43", align 8
%"50" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"50", ptr addrspace(5) %"44", align 8
%"52" = load i64, ptr addrspace(5) %"43", align 8
%"66" = inttoptr i64 %"52" to ptr
%"51" = load i64, ptr %"66", align 4
store i64 %"51", ptr addrspace(5) %"45", align 4
%"53" = load i64, ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr %"66", align 8
store i64 %"51", ptr addrspace(5) %"45", align 8
%"53" = load i64, ptr addrspace(5) %"43", align 8
%"67" = inttoptr i64 %"53" to ptr
%"37" = getelementptr inbounds i8, ptr %"67", i64 8
%"54" = load i64, ptr %"37", align 4
store i64 %"54", ptr addrspace(5) %"46", align 4
%"56" = load i64, ptr addrspace(5) %"45", align 4
%"57" = load i64, ptr addrspace(5) %"46", align 4
%"54" = load i64, ptr %"37", align 8
store i64 %"54", ptr addrspace(5) %"46", align 8
%"56" = load i64, ptr addrspace(5) %"45", align 8
%"57" = load i64, ptr addrspace(5) %"46", align 8
%"55" = icmp ult i64 %"56", %"57"
store i1 %"55", ptr addrspace(5) %"48", align 1
%"59" = load i1, ptr addrspace(5) %"48", align 1
@ -35,7 +35,7 @@ define amdgpu_kernel void @pred_not(ptr addrspace(4) byref(i64) %"41", ptr addrs
br i1 %"60", label %"16", label %"17"
"16": ; preds = %"40"
store i64 1, ptr addrspace(5) %"47", align 4
store i64 1, ptr addrspace(5) %"47", align 8
br label %"17"
"17": ; preds = %"16", %"40"
@ -43,14 +43,14 @@ define amdgpu_kernel void @pred_not(ptr addrspace(4) byref(i64) %"41", ptr addrs
br i1 %"62", label %"19", label %"18"
"18": ; preds = %"17"
store i64 2, ptr addrspace(5) %"47", align 4
store i64 2, ptr addrspace(5) %"47", align 8
br label %"19"
"19": ; preds = %"18", %"17"
%"64" = load i64, ptr addrspace(5) %"44", align 4
%"65" = load i64, ptr addrspace(5) %"47", align 4
%"64" = load i64, ptr addrspace(5) %"44", align 8
%"65" = load i64, ptr addrspace(5) %"47", align 8
%"68" = inttoptr i64 %"64" to ptr
store i64 %"65", ptr %"68", align 4
store i64 %"65", ptr %"68", align 8
ret void
}

14
ptx/src/test/ll/prmt.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @prmt(ptr addrspace(4) byref(i64) %"33", ptr addrspace
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load i32, ptr %"31", align 4
@ -28,7 +28,7 @@ define amdgpu_kernel void @prmt(ptr addrspace(4) byref(i64) %"33", ptr addrspace
%3 = bitcast i32 %"47" to <4 x i8>
%"52" = shufflevector <4 x i8> %2, <4 x i8> %3, <4 x i32> <i32 4, i32 0, i32 6, i32 7>
store <4 x i8> %"52", ptr addrspace(5) %"38", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load i32, ptr addrspace(5) %"38", align 4
%"55" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"55", align 4

12
ptx/src/test/ll/rcp.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @rcp(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load float, ptr %"43", align 4
store float %"37", ptr addrspace(5) %"34", align 4
%"40" = load float, ptr addrspace(5) %"34", align 4
%"39" = call float @llvm.amdgcn.rcp.f32(float %"40")
store float %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load float, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store float %"42", ptr %"44", align 4

28
ptx/src/test/ll/reg_local.ll
View File

@ -9,27 +9,27 @@ define amdgpu_kernel void @reg_local(ptr addrspace(4) byref(i64) %"37", ptr addr
br label %"36"
"36": ; preds = %1
%"42" = load i64, ptr addrspace(4) %"37", align 4
store i64 %"42", ptr addrspace(5) %"39", align 4
%"43" = load i64, ptr addrspace(4) %"38", align 4
store i64 %"43", ptr addrspace(5) %"40", align 4
%"45" = load i64, ptr addrspace(5) %"39", align 4
%"42" = load i64, ptr addrspace(4) %"37", align 8
store i64 %"42", ptr addrspace(5) %"39", align 8
%"43" = load i64, ptr addrspace(4) %"38", align 8
store i64 %"43", ptr addrspace(5) %"40", align 8
%"45" = load i64, ptr addrspace(5) %"39", align 8
%"51" = inttoptr i64 %"45" to ptr addrspace(1)
%"50" = load i64, ptr addrspace(1) %"51", align 4
store i64 %"50", ptr addrspace(5) %"41", align 4
%"46" = load i64, ptr addrspace(5) %"41", align 4
%"50" = load i64, ptr addrspace(1) %"51", align 8
store i64 %"50", ptr addrspace(5) %"41", align 8
%"46" = load i64, ptr addrspace(5) %"41", align 8
%"31" = add i64 %"46", 1
%"52" = addrspacecast ptr addrspace(5) %"10" to ptr
store i64 %"31", ptr %"52", align 4
store i64 %"31", ptr %"52", align 8
%"54" = addrspacecast ptr addrspace(5) %"10" to ptr
%"33" = getelementptr inbounds i8, ptr %"54", i64 0
%"55" = load i64, ptr %"33", align 4
store i64 %"55", ptr addrspace(5) %"41", align 4
%"48" = load i64, ptr addrspace(5) %"40", align 4
%"55" = load i64, ptr %"33", align 8
store i64 %"55", ptr addrspace(5) %"41", align 8
%"48" = load i64, ptr addrspace(5) %"40", align 8
%"56" = inttoptr i64 %"48" to ptr addrspace(1)
%"35" = getelementptr inbounds i8, ptr addrspace(1) %"56", i64 0
%"49" = load i64, ptr addrspace(5) %"41", align 4
store i64 %"49", ptr addrspace(1) %"35", align 4
%"49" = load i64, ptr addrspace(5) %"41", align 8
store i64 %"49", ptr addrspace(1) %"35", align 8
ret void
}

14
ptx/src/test/ll/rem.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @rem(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
br label %"32"
"32": ; preds = %1
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"35", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"35", align 8
%"50" = inttoptr i64 %"42" to ptr
%"41" = load i32, ptr %"50", align 4
store i32 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 4
%"43" = load i64, ptr addrspace(5) %"35", align 8
%"51" = inttoptr i64 %"43" to ptr
%"31" = getelementptr inbounds i8, ptr %"51", i64 4
%"44" = load i32, ptr %"31", align 4
@ -26,7 +26,7 @@ define amdgpu_kernel void @rem(ptr addrspace(4) byref(i64) %"33", ptr addrspace(
%"47" = load i32, ptr addrspace(5) %"38", align 4
%"45" = srem i32 %"46", %"47"
store i32 %"45", ptr addrspace(5) %"37", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 4
%"48" = load i64, ptr addrspace(5) %"36", align 8
%"49" = load i32, ptr addrspace(5) %"37", align 4
%"52" = inttoptr i64 %"48" to ptr
store i32 %"49", ptr %"52", align 4

12
ptx/src/test/ll/rsqrt.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @rsqrt(ptr addrspace(4) byref(i64) %"30", ptr addrspac
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load double, ptr %"43", align 8
store double %"37", ptr addrspace(5) %"34", align 8
%"40" = load double, ptr addrspace(5) %"34", align 8
%"39" = call double @llvm.amdgcn.rsq.f64(double %"40")
store double %"39", ptr addrspace(5) %"34", align 8
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load double, ptr addrspace(5) %"34", align 8
%"44" = inttoptr i64 %"41" to ptr
store double %"42", ptr %"44", align 8

14
ptx/src/test/ll/selp.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @selp(ptr addrspace(4) byref(i64) %"34", ptr addrspace
br label %"33"
"33": ; preds = %1
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"51" = inttoptr i64 %"43" to ptr
%"42" = load i16, ptr %"51", align 2
store i16 %"42", ptr addrspace(5) %"38", align 2
%"44" = load i64, ptr addrspace(5) %"36", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"52" = inttoptr i64 %"44" to ptr
%"31" = getelementptr inbounds i8, ptr %"52", i64 2
%"45" = load i16, ptr %"31", align 2
@ -26,7 +26,7 @@ define amdgpu_kernel void @selp(ptr addrspace(4) byref(i64) %"34", ptr addrspace
%"48" = load i16, ptr addrspace(5) %"39", align 2
%"46" = select i1 false, i16 %"47", i16 %"48"
store i16 %"46", ptr addrspace(5) %"38", align 2
%"49" = load i64, ptr addrspace(5) %"37", align 4
%"49" = load i64, ptr addrspace(5) %"37", align 8
%"50" = load i16, ptr addrspace(5) %"38", align 2
%"53" = inttoptr i64 %"49" to ptr
store i16 %"50", ptr %"53", align 2

14
ptx/src/test/ll/selp_true.ll
View File

@ -9,15 +9,15 @@ define amdgpu_kernel void @selp_true(ptr addrspace(4) byref(i64) %"34", ptr addr
br label %"33"
"33": ; preds = %1
%"40" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"41" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"41", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"41" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"41", ptr addrspace(5) %"37", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%"51" = inttoptr i64 %"43" to ptr
%"42" = load i16, ptr %"51", align 2
store i16 %"42", ptr addrspace(5) %"38", align 2
%"44" = load i64, ptr addrspace(5) %"36", align 4
%"44" = load i64, ptr addrspace(5) %"36", align 8
%"52" = inttoptr i64 %"44" to ptr
%"31" = getelementptr inbounds i8, ptr %"52", i64 2
%"45" = load i16, ptr %"31", align 2
@ -26,7 +26,7 @@ define amdgpu_kernel void @selp_true(ptr addrspace(4) byref(i64) %"34", ptr addr
%"48" = load i16, ptr addrspace(5) %"39", align 2
%"46" = select i1 true, i16 %"47", i16 %"48"
store i16 %"46", ptr addrspace(5) %"38", align 2
%"49" = load i64, ptr addrspace(5) %"37", align 4
%"49" = load i64, ptr addrspace(5) %"37", align 8
%"50" = load i16, ptr addrspace(5) %"38", align 2
%"53" = inttoptr i64 %"49" to ptr
store i16 %"50", ptr %"53", align 2

34
ptx/src/test/ll/setp.ll
View File

@ -11,28 +11,28 @@ define amdgpu_kernel void @setp(ptr addrspace(4) byref(i64) %"41", ptr addrspace
br label %"40"
"40": ; preds = %1
%"49" = load i64, ptr addrspace(4) %"41", align 4
store i64 %"49", ptr addrspace(5) %"43", align 4
%"50" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"50", ptr addrspace(5) %"44", align 4
%"52" = load i64, ptr addrspace(5) %"43", align 4
%"49" = load i64, ptr addrspace(4) %"41", align 8
store i64 %"49", ptr addrspace(5) %"43", align 8
%"50" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"50", ptr addrspace(5) %"44", align 8
%"52" = load i64, ptr addrspace(5) %"43", align 8
%"64" = inttoptr i64 %"52" to ptr
%"51" = load i64, ptr %"64", align 4
store i64 %"51", ptr addrspace(5) %"45", align 4
%"53" = load i64, ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr %"64", align 8
store i64 %"51", ptr addrspace(5) %"45", align 8
%"53" = load i64, ptr addrspace(5) %"43", align 8
%"65" = inttoptr i64 %"53" to ptr
%"37" = getelementptr inbounds i8, ptr %"65", i64 8
%"54" = load i64, ptr %"37", align 4
store i64 %"54", ptr addrspace(5) %"46", align 4
%"56" = load i64, ptr addrspace(5) %"45", align 4
%"57" = load i64, ptr addrspace(5) %"46", align 4
%"54" = load i64, ptr %"37", align 8
store i64 %"54", ptr addrspace(5) %"46", align 8
%"56" = load i64, ptr addrspace(5) %"45", align 8
%"57" = load i64, ptr addrspace(5) %"46", align 8
%"55" = icmp ult i64 %"56", %"57"
store i1 %"55", ptr addrspace(5) %"48", align 1
%"58" = load i1, ptr addrspace(5) %"48", align 1
br i1 %"58", label %"16", label %"17"
"16": ; preds = %"40"
store i64 1, ptr addrspace(5) %"47", align 4
store i64 1, ptr addrspace(5) %"47", align 8
br label %"17"
"17": ; preds = %"16", %"40"
@ -40,14 +40,14 @@ define amdgpu_kernel void @setp(ptr addrspace(4) byref(i64) %"41", ptr addrspace
br i1 %"60", label %"19", label %"18"
"18": ; preds = %"17"
store i64 2, ptr addrspace(5) %"47", align 4
store i64 2, ptr addrspace(5) %"47", align 8
br label %"19"
"19": ; preds = %"18", %"17"
%"62" = load i64, ptr addrspace(5) %"44", align 4
%"63" = load i64, ptr addrspace(5) %"47", align 4
%"62" = load i64, ptr addrspace(5) %"44", align 8
%"63" = load i64, ptr addrspace(5) %"47", align 8
%"66" = inttoptr i64 %"62" to ptr
store i64 %"63", ptr %"66", align 4
store i64 %"63", ptr %"66", align 8
ret void
}

14
ptx/src/test/ll/setp_gt.ll
View File

@ -11,15 +11,15 @@ define amdgpu_kernel void @setp_gt(ptr addrspace(4) byref(i64) %"39", ptr addrsp
br label %"38"
"38": ; preds = %1
%"47" = load i64, ptr addrspace(4) %"39", align 4
store i64 %"47", ptr addrspace(5) %"41", align 4
%"48" = load i64, ptr addrspace(4) %"40", align 4
store i64 %"48", ptr addrspace(5) %"42", align 4
%"50" = load i64, ptr addrspace(5) %"41", align 4
%"47" = load i64, ptr addrspace(4) %"39", align 8
store i64 %"47", ptr addrspace(5) %"41", align 8
%"48" = load i64, ptr addrspace(4) %"40", align 8
store i64 %"48", ptr addrspace(5) %"42", align 8
%"50" = load i64, ptr addrspace(5) %"41", align 8
%"64" = inttoptr i64 %"50" to ptr
%"49" = load float, ptr %"64", align 4
store float %"49", ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr addrspace(5) %"41", align 4
%"51" = load i64, ptr addrspace(5) %"41", align 8
%"65" = inttoptr i64 %"51" to ptr
%"37" = getelementptr inbounds i8, ptr %"65", i64 4
%"52" = load float, ptr %"37", align 4
@ -46,7 +46,7 @@ define amdgpu_kernel void @setp_gt(ptr addrspace(4) byref(i64) %"39", ptr addrsp
br label %"19"
"19": ; preds = %"18", %"17"
%"62" = load i64, ptr addrspace(5) %"42", align 4
%"62" = load i64, ptr addrspace(5) %"42", align 8
%"63" = load float, ptr addrspace(5) %"45", align 4
%"66" = inttoptr i64 %"62" to ptr
store float %"63", ptr %"66", align 4

14
ptx/src/test/ll/setp_leu.ll
View File

@ -11,15 +11,15 @@ define amdgpu_kernel void @setp_leu(ptr addrspace(4) byref(i64) %"39", ptr addrs
br label %"38"
"38": ; preds = %1
%"47" = load i64, ptr addrspace(4) %"39", align 4
store i64 %"47", ptr addrspace(5) %"41", align 4
%"48" = load i64, ptr addrspace(4) %"40", align 4
store i64 %"48", ptr addrspace(5) %"42", align 4
%"50" = load i64, ptr addrspace(5) %"41", align 4
%"47" = load i64, ptr addrspace(4) %"39", align 8
store i64 %"47", ptr addrspace(5) %"41", align 8
%"48" = load i64, ptr addrspace(4) %"40", align 8
store i64 %"48", ptr addrspace(5) %"42", align 8
%"50" = load i64, ptr addrspace(5) %"41", align 8
%"64" = inttoptr i64 %"50" to ptr
%"49" = load float, ptr %"64", align 4
store float %"49", ptr addrspace(5) %"43", align 4
%"51" = load i64, ptr addrspace(5) %"41", align 4
%"51" = load i64, ptr addrspace(5) %"41", align 8
%"65" = inttoptr i64 %"51" to ptr
%"37" = getelementptr inbounds i8, ptr %"65", i64 4
%"52" = load float, ptr %"37", align 4
@ -46,7 +46,7 @@ define amdgpu_kernel void @setp_leu(ptr addrspace(4) byref(i64) %"39", ptr addrs
br label %"19"
"19": ; preds = %"18", %"17"
%"62" = load i64, ptr addrspace(5) %"42", align 4
%"62" = load i64, ptr addrspace(5) %"42", align 8
%"63" = load float, ptr addrspace(5) %"45", align 4
%"66" = inttoptr i64 %"62" to ptr
store float %"63", ptr %"66", align 4

32
ptx/src/test/ll/setp_nan.ll
View File

@ -17,45 +17,45 @@ define amdgpu_kernel void @setp_nan(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"82"
"82": ; preds = %1
%"97" = load i64, ptr addrspace(4) %"83", align 4
store i64 %"97", ptr addrspace(5) %"85", align 4
%"98" = load i64, ptr addrspace(4) %"84", align 4
store i64 %"98", ptr addrspace(5) %"86", align 4
%"100" = load i64, ptr addrspace(5) %"85", align 4
%"97" = load i64, ptr addrspace(4) %"83", align 8
store i64 %"97", ptr addrspace(5) %"85", align 8
%"98" = load i64, ptr addrspace(4) %"84", align 8
store i64 %"98", ptr addrspace(5) %"86", align 8
%"100" = load i64, ptr addrspace(5) %"85", align 8
%"151" = inttoptr i64 %"100" to ptr
%"99" = load float, ptr %"151", align 4
store float %"99", ptr addrspace(5) %"87", align 4
%"101" = load i64, ptr addrspace(5) %"85", align 4
%"101" = load i64, ptr addrspace(5) %"85", align 8
%"152" = inttoptr i64 %"101" to ptr
%"55" = getelementptr inbounds i8, ptr %"152", i64 4
%"102" = load float, ptr %"55", align 4
store float %"102", ptr addrspace(5) %"88", align 4
%"103" = load i64, ptr addrspace(5) %"85", align 4
%"103" = load i64, ptr addrspace(5) %"85", align 8
%"153" = inttoptr i64 %"103" to ptr
%"57" = getelementptr inbounds i8, ptr %"153", i64 8
%"104" = load float, ptr %"57", align 4
store float %"104", ptr addrspace(5) %"89", align 4
%"105" = load i64, ptr addrspace(5) %"85", align 4
%"105" = load i64, ptr addrspace(5) %"85", align 8
%"154" = inttoptr i64 %"105" to ptr
%"59" = getelementptr inbounds i8, ptr %"154", i64 12
%"106" = load float, ptr %"59", align 4
store float %"106", ptr addrspace(5) %"90", align 4
%"107" = load i64, ptr addrspace(5) %"85", align 4
%"107" = load i64, ptr addrspace(5) %"85", align 8
%"155" = inttoptr i64 %"107" to ptr
%"61" = getelementptr inbounds i8, ptr %"155", i64 16
%"108" = load float, ptr %"61", align 4
store float %"108", ptr addrspace(5) %"91", align 4
%"109" = load i64, ptr addrspace(5) %"85", align 4
%"109" = load i64, ptr addrspace(5) %"85", align 8
%"156" = inttoptr i64 %"109" to ptr
%"63" = getelementptr inbounds i8, ptr %"156", i64 20
%"110" = load float, ptr %"63", align 4
store float %"110", ptr addrspace(5) %"92", align 4
%"111" = load i64, ptr addrspace(5) %"85", align 4
%"111" = load i64, ptr addrspace(5) %"85", align 8
%"157" = inttoptr i64 %"111" to ptr
%"65" = getelementptr inbounds i8, ptr %"157", i64 24
%"112" = load float, ptr %"65", align 4
store float %"112", ptr addrspace(5) %"93", align 4
%"113" = load i64, ptr addrspace(5) %"85", align 4
%"113" = load i64, ptr addrspace(5) %"85", align 8
%"158" = inttoptr i64 %"113" to ptr
%"67" = getelementptr inbounds i8, ptr %"158", i64 28
%"114" = load float, ptr %"67", align 4
@ -80,7 +80,7 @@ define amdgpu_kernel void @setp_nan(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"25"
"25": ; preds = %"24", %"23"
%"122" = load i64, ptr addrspace(5) %"86", align 4
%"122" = load i64, ptr addrspace(5) %"86", align 8
%"123" = load i32, ptr addrspace(5) %"95", align 4
%"159" = inttoptr i64 %"122" to ptr
store i32 %"123", ptr %"159", align 4
@ -104,7 +104,7 @@ define amdgpu_kernel void @setp_nan(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"29"
"29": ; preds = %"28", %"27"
%"131" = load i64, ptr addrspace(5) %"86", align 4
%"131" = load i64, ptr addrspace(5) %"86", align 8
%"160" = inttoptr i64 %"131" to ptr
%"73" = getelementptr inbounds i8, ptr %"160", i64 4
%"132" = load i32, ptr addrspace(5) %"95", align 4
@ -129,7 +129,7 @@ define amdgpu_kernel void @setp_nan(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"33"
"33": ; preds = %"32", %"31"
%"140" = load i64, ptr addrspace(5) %"86", align 4
%"140" = load i64, ptr addrspace(5) %"86", align 8
%"161" = inttoptr i64 %"140" to ptr
%"77" = getelementptr inbounds i8, ptr %"161", i64 8
%"141" = load i32, ptr addrspace(5) %"95", align 4
@ -154,7 +154,7 @@ define amdgpu_kernel void @setp_nan(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"37"
"37": ; preds = %"36", %"35"
%"149" = load i64, ptr addrspace(5) %"86", align 4
%"149" = load i64, ptr addrspace(5) %"86", align 8
%"162" = inttoptr i64 %"149" to ptr
%"81" = getelementptr inbounds i8, ptr %"162", i64 12
%"150" = load i32, ptr addrspace(5) %"95", align 4

32
ptx/src/test/ll/setp_num.ll
View File

@ -17,45 +17,45 @@ define amdgpu_kernel void @setp_num(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"82"
"82": ; preds = %1
%"97" = load i64, ptr addrspace(4) %"83", align 4
store i64 %"97", ptr addrspace(5) %"85", align 4
%"98" = load i64, ptr addrspace(4) %"84", align 4
store i64 %"98", ptr addrspace(5) %"86", align 4
%"100" = load i64, ptr addrspace(5) %"85", align 4
%"97" = load i64, ptr addrspace(4) %"83", align 8
store i64 %"97", ptr addrspace(5) %"85", align 8
%"98" = load i64, ptr addrspace(4) %"84", align 8
store i64 %"98", ptr addrspace(5) %"86", align 8
%"100" = load i64, ptr addrspace(5) %"85", align 8
%"151" = inttoptr i64 %"100" to ptr
%"99" = load float, ptr %"151", align 4
store float %"99", ptr addrspace(5) %"87", align 4
%"101" = load i64, ptr addrspace(5) %"85", align 4
%"101" = load i64, ptr addrspace(5) %"85", align 8
%"152" = inttoptr i64 %"101" to ptr
%"55" = getelementptr inbounds i8, ptr %"152", i64 4
%"102" = load float, ptr %"55", align 4
store float %"102", ptr addrspace(5) %"88", align 4
%"103" = load i64, ptr addrspace(5) %"85", align 4
%"103" = load i64, ptr addrspace(5) %"85", align 8
%"153" = inttoptr i64 %"103" to ptr
%"57" = getelementptr inbounds i8, ptr %"153", i64 8
%"104" = load float, ptr %"57", align 4
store float %"104", ptr addrspace(5) %"89", align 4
%"105" = load i64, ptr addrspace(5) %"85", align 4
%"105" = load i64, ptr addrspace(5) %"85", align 8
%"154" = inttoptr i64 %"105" to ptr
%"59" = getelementptr inbounds i8, ptr %"154", i64 12
%"106" = load float, ptr %"59", align 4
store float %"106", ptr addrspace(5) %"90", align 4
%"107" = load i64, ptr addrspace(5) %"85", align 4
%"107" = load i64, ptr addrspace(5) %"85", align 8
%"155" = inttoptr i64 %"107" to ptr
%"61" = getelementptr inbounds i8, ptr %"155", i64 16
%"108" = load float, ptr %"61", align 4
store float %"108", ptr addrspace(5) %"91", align 4
%"109" = load i64, ptr addrspace(5) %"85", align 4
%"109" = load i64, ptr addrspace(5) %"85", align 8
%"156" = inttoptr i64 %"109" to ptr
%"63" = getelementptr inbounds i8, ptr %"156", i64 20
%"110" = load float, ptr %"63", align 4
store float %"110", ptr addrspace(5) %"92", align 4
%"111" = load i64, ptr addrspace(5) %"85", align 4
%"111" = load i64, ptr addrspace(5) %"85", align 8
%"157" = inttoptr i64 %"111" to ptr
%"65" = getelementptr inbounds i8, ptr %"157", i64 24
%"112" = load float, ptr %"65", align 4
store float %"112", ptr addrspace(5) %"93", align 4
%"113" = load i64, ptr addrspace(5) %"85", align 4
%"113" = load i64, ptr addrspace(5) %"85", align 8
%"158" = inttoptr i64 %"113" to ptr
%"67" = getelementptr inbounds i8, ptr %"158", i64 28
%"114" = load float, ptr %"67", align 4
@ -80,7 +80,7 @@ define amdgpu_kernel void @setp_num(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"25"
"25": ; preds = %"24", %"23"
%"122" = load i64, ptr addrspace(5) %"86", align 4
%"122" = load i64, ptr addrspace(5) %"86", align 8
%"123" = load i32, ptr addrspace(5) %"95", align 4
%"159" = inttoptr i64 %"122" to ptr
store i32 %"123", ptr %"159", align 4
@ -104,7 +104,7 @@ define amdgpu_kernel void @setp_num(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"29"
"29": ; preds = %"28", %"27"
%"131" = load i64, ptr addrspace(5) %"86", align 4
%"131" = load i64, ptr addrspace(5) %"86", align 8
%"160" = inttoptr i64 %"131" to ptr
%"73" = getelementptr inbounds i8, ptr %"160", i64 4
%"132" = load i32, ptr addrspace(5) %"95", align 4
@ -129,7 +129,7 @@ define amdgpu_kernel void @setp_num(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"33"
"33": ; preds = %"32", %"31"
%"140" = load i64, ptr addrspace(5) %"86", align 4
%"140" = load i64, ptr addrspace(5) %"86", align 8
%"161" = inttoptr i64 %"140" to ptr
%"77" = getelementptr inbounds i8, ptr %"161", i64 8
%"141" = load i32, ptr addrspace(5) %"95", align 4
@ -154,7 +154,7 @@ define amdgpu_kernel void @setp_num(ptr addrspace(4) byref(i64) %"83", ptr addrs
br label %"37"
"37": ; preds = %"36", %"35"
%"149" = load i64, ptr addrspace(5) %"86", align 4
%"149" = load i64, ptr addrspace(5) %"86", align 8
%"162" = inttoptr i64 %"149" to ptr
%"81" = getelementptr inbounds i8, ptr %"162", i64 12
%"150" = load i32, ptr addrspace(5) %"95", align 4

28
ptx/src/test/ll/shared_ptr_32.ll
View File

@ -12,28 +12,28 @@ define amdgpu_kernel void @shared_ptr_32(ptr addrspace(4) byref(i64) %"35", ptr
br label %"34"
"34": ; preds = %1
%"42" = load i64, ptr addrspace(4) %"35", align 4
store i64 %"42", ptr addrspace(5) %"37", align 4
%"43" = load i64, ptr addrspace(4) %"36", align 4
store i64 %"43", ptr addrspace(5) %"38", align 4
%"42" = load i64, ptr addrspace(4) %"35", align 8
store i64 %"42", ptr addrspace(5) %"37", align 8
%"43" = load i64, ptr addrspace(4) %"36", align 8
store i64 %"43", ptr addrspace(5) %"38", align 8
store i32 ptrtoint (ptr addrspace(3) @shared_mem1 to i32), ptr addrspace(5) %"39", align 4
%"46" = load i64, ptr addrspace(5) %"37", align 4
%"46" = load i64, ptr addrspace(5) %"37", align 8
%"54" = inttoptr i64 %"46" to ptr addrspace(1)
%"45" = load i64, ptr addrspace(1) %"54", align 4
store i64 %"45", ptr addrspace(5) %"40", align 4
%"45" = load i64, ptr addrspace(1) %"54", align 8
store i64 %"45", ptr addrspace(5) %"40", align 8
%"47" = load i32, ptr addrspace(5) %"39", align 4
%"48" = load i64, ptr addrspace(5) %"40", align 4
%"48" = load i64, ptr addrspace(5) %"40", align 8
%"55" = inttoptr i32 %"47" to ptr addrspace(3)
store i64 %"48", ptr addrspace(3) %"55", align 4
store i64 %"48", ptr addrspace(3) %"55", align 8
%"49" = load i32, ptr addrspace(5) %"39", align 4
%"56" = inttoptr i32 %"49" to ptr addrspace(3)
%"33" = getelementptr inbounds i8, ptr addrspace(3) %"56", i64 0
%"50" = load i64, ptr addrspace(3) %"33", align 4
store i64 %"50", ptr addrspace(5) %"41", align 4
%"51" = load i64, ptr addrspace(5) %"38", align 4
%"52" = load i64, ptr addrspace(5) %"41", align 4
%"50" = load i64, ptr addrspace(3) %"33", align 8
store i64 %"50", ptr addrspace(5) %"41", align 8
%"51" = load i64, ptr addrspace(5) %"38", align 8
%"52" = load i64, ptr addrspace(5) %"41", align 8
%"57" = inttoptr i64 %"51" to ptr addrspace(1)
store i64 %"52", ptr addrspace(1) %"57", align 4
store i64 %"52", ptr addrspace(1) %"57", align 8
ret void
}

34
ptx/src/test/ll/shared_ptr_take_address.ll
View File

@ -12,27 +12,27 @@ define amdgpu_kernel void @shared_ptr_take_address(ptr addrspace(4) byref(i64) %
br label %"32"
"32": ; preds = %1
%"40" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"40", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(4) %"34", align 4
store i64 %"41", ptr addrspace(5) %"36", align 4
store i64 ptrtoint (ptr addrspace(3) @shared_mem to i64), ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"40" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"40", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(4) %"34", align 8
store i64 %"41", ptr addrspace(5) %"36", align 8
store i64 ptrtoint (ptr addrspace(3) @shared_mem to i64), ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"52" = inttoptr i64 %"44" to ptr addrspace(1)
%"43" = load i64, ptr addrspace(1) %"52", align 4
store i64 %"43", ptr addrspace(5) %"38", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
%"46" = load i64, ptr addrspace(5) %"38", align 4
%"43" = load i64, ptr addrspace(1) %"52", align 8
store i64 %"43", ptr addrspace(5) %"38", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"46" = load i64, ptr addrspace(5) %"38", align 8
%"53" = inttoptr i64 %"45" to ptr addrspace(3)
store i64 %"46", ptr addrspace(3) %"53", align 4
%"48" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"46", ptr addrspace(3) %"53", align 8
%"48" = load i64, ptr addrspace(5) %"37", align 8
%"54" = inttoptr i64 %"48" to ptr addrspace(3)
%"47" = load i64, ptr addrspace(3) %"54", align 4
store i64 %"47", ptr addrspace(5) %"39", align 4
%"49" = load i64, ptr addrspace(5) %"36", align 4
%"50" = load i64, ptr addrspace(5) %"39", align 4
%"47" = load i64, ptr addrspace(3) %"54", align 8
store i64 %"47", ptr addrspace(5) %"39", align 8
%"49" = load i64, ptr addrspace(5) %"36", align 8
%"50" = load i64, ptr addrspace(5) %"39", align 8
%"55" = inttoptr i64 %"49" to ptr addrspace(1)
store i64 %"50", ptr addrspace(1) %"55", align 4
store i64 %"50", ptr addrspace(1) %"55", align 8
ret void
}

56
ptx/src/test/ll/shared_unify_extern.ll
View File

@ -11,15 +11,15 @@ define i64 @add() #0 {
br label %"41"
"41": ; preds = %1
%"49" = load i64, ptr addrspace(3) @shared_mod, align 4
store i64 %"49", ptr addrspace(5) %"47", align 4
%"50" = load i64, ptr addrspace(3) @shared_ex, align 4
store i64 %"50", ptr addrspace(5) %"48", align 4
%"52" = load i64, ptr addrspace(5) %"48", align 4
%"53" = load i64, ptr addrspace(5) %"47", align 4
%"49" = load i64, ptr addrspace(3) @shared_mod, align 8
store i64 %"49", ptr addrspace(5) %"47", align 8
%"50" = load i64, ptr addrspace(3) @shared_ex, align 8
store i64 %"50", ptr addrspace(5) %"48", align 8
%"52" = load i64, ptr addrspace(5) %"48", align 8
%"53" = load i64, ptr addrspace(5) %"47", align 8
%"75" = add i64 %"52", %"53"
store i64 %"75", ptr addrspace(5) %"46", align 4
%2 = load i64, ptr addrspace(5) %"46", align 4
store i64 %"75", ptr addrspace(5) %"46", align 8
%2 = load i64, ptr addrspace(5) %"46", align 8
ret i64 %2
}
@ -31,13 +31,13 @@ define i64 @set_shared_temp1(i64 %"15") #0 {
br label %"42"
"42": ; preds = %1
store i64 %"15", ptr addrspace(3) @shared_ex, align 4
store i64 %"15", ptr addrspace(3) @shared_ex, align 8
%"55" = call i64 @add()
store i64 %"55", ptr addrspace(5) %"54", align 4
store i64 %"55", ptr addrspace(5) %"54", align 8
br label %"43"
"43": ; preds = %"42"
%2 = load i64, ptr addrspace(5) %"54", align 4
%2 = load i64, ptr addrspace(5) %"54", align 8
ret i64 %2
}
@ -52,31 +52,31 @@ define amdgpu_kernel void @shared_unify_extern(ptr addrspace(4) byref(i64) %"56"
br label %"44"
"44": ; preds = %1
%"62" = load i64, ptr addrspace(4) %"56", align 4
store i64 %"62", ptr addrspace(5) %"58", align 4
%"63" = load i64, ptr addrspace(4) %"57", align 4
store i64 %"63", ptr addrspace(5) %"59", align 4
%"65" = load i64, ptr addrspace(5) %"58", align 4
%"62" = load i64, ptr addrspace(4) %"56", align 8
store i64 %"62", ptr addrspace(5) %"58", align 8
%"63" = load i64, ptr addrspace(4) %"57", align 8
store i64 %"63", ptr addrspace(5) %"59", align 8
%"65" = load i64, ptr addrspace(5) %"58", align 8
%"78" = inttoptr i64 %"65" to ptr addrspace(1)
%"64" = load i64, ptr addrspace(1) %"78", align 4
store i64 %"64", ptr addrspace(5) %"60", align 4
%"66" = load i64, ptr addrspace(5) %"58", align 4
%"64" = load i64, ptr addrspace(1) %"78", align 8
store i64 %"64", ptr addrspace(5) %"60", align 8
%"66" = load i64, ptr addrspace(5) %"58", align 8
%"79" = inttoptr i64 %"66" to ptr addrspace(1)
%"40" = getelementptr inbounds i8, ptr addrspace(1) %"79", i64 8
%"67" = load i64, ptr addrspace(1) %"40", align 4
store i64 %"67", ptr addrspace(5) %"61", align 4
%"68" = load i64, ptr addrspace(5) %"61", align 4
store i64 %"68", ptr addrspace(3) @shared_mod, align 4
%"70" = load i64, ptr addrspace(5) %"60", align 4
%"67" = load i64, ptr addrspace(1) %"40", align 8
store i64 %"67", ptr addrspace(5) %"61", align 8
%"68" = load i64, ptr addrspace(5) %"61", align 8
store i64 %"68", ptr addrspace(3) @shared_mod, align 8
%"70" = load i64, ptr addrspace(5) %"60", align 8
%"81" = call i64 @set_shared_temp1(i64 %"70")
store i64 %"81", ptr addrspace(5) %"61", align 4
store i64 %"81", ptr addrspace(5) %"61", align 8
br label %"45"
"45": ; preds = %"44"
%"71" = load i64, ptr addrspace(5) %"59", align 4
%"72" = load i64, ptr addrspace(5) %"61", align 4
%"71" = load i64, ptr addrspace(5) %"59", align 8
%"72" = load i64, ptr addrspace(5) %"61", align 8
%"83" = inttoptr i64 %"71" to ptr
store i64 %"72", ptr %"83", align 4
store i64 %"72", ptr %"83", align 8
ret void
}

52
ptx/src/test/ll/shared_unify_local.ll
View File

@ -10,14 +10,14 @@ define i64 @add(i64 %"10") #0 {
br label %"42"
"42": ; preds = %1
store i64 %"10", ptr addrspace(3) @shared_mod, align 4
%"49" = load i64, ptr addrspace(3) @shared_mod, align 4
store i64 %"49", ptr addrspace(5) %"48", align 4
%"101" = load i64, ptr addrspace(3) @shared_ex, align 4
%"51" = load i64, ptr addrspace(5) %"48", align 4
store i64 %"10", ptr addrspace(3) @shared_mod, align 8
%"49" = load i64, ptr addrspace(3) @shared_mod, align 8
store i64 %"49", ptr addrspace(5) %"48", align 8
%"101" = load i64, ptr addrspace(3) @shared_ex, align 8
%"51" = load i64, ptr addrspace(5) %"48", align 8
%"72" = add i64 %"101", %"51"
store i64 %"72", ptr addrspace(5) %"47", align 4
%2 = load i64, ptr addrspace(5) %"47", align 4
store i64 %"72", ptr addrspace(5) %"47", align 8
%2 = load i64, ptr addrspace(5) %"47", align 8
ret i64 %2
}
@ -29,13 +29,13 @@ define i64 @set_shared_temp1(i64 %"15", i64 %"16") #0 {
br label %"43"
"43": ; preds = %1
store i64 %"15", ptr addrspace(3) @shared_ex, align 4
store i64 %"15", ptr addrspace(3) @shared_ex, align 8
%"53" = call i64 @add(i64 %"16")
store i64 %"53", ptr addrspace(5) %"52", align 4
store i64 %"53", ptr addrspace(5) %"52", align 8
br label %"44"
"44": ; preds = %"43"
%2 = load i64, ptr addrspace(5) %"52", align 4
%2 = load i64, ptr addrspace(5) %"52", align 8
ret i64 %2
}
@ -50,30 +50,30 @@ define amdgpu_kernel void @shared_unify_local(ptr addrspace(4) byref(i64) %"54",
br label %"45"
"45": ; preds = %1
%"60" = load i64, ptr addrspace(4) %"54", align 4
store i64 %"60", ptr addrspace(5) %"56", align 4
%"61" = load i64, ptr addrspace(4) %"55", align 4
store i64 %"61", ptr addrspace(5) %"57", align 4
%"63" = load i64, ptr addrspace(5) %"56", align 4
%"60" = load i64, ptr addrspace(4) %"54", align 8
store i64 %"60", ptr addrspace(5) %"56", align 8
%"61" = load i64, ptr addrspace(4) %"55", align 8
store i64 %"61", ptr addrspace(5) %"57", align 8
%"63" = load i64, ptr addrspace(5) %"56", align 8
%"75" = inttoptr i64 %"63" to ptr addrspace(1)
%"62" = load i64, ptr addrspace(1) %"75", align 4
store i64 %"62", ptr addrspace(5) %"58", align 4
%"64" = load i64, ptr addrspace(5) %"56", align 4
%"62" = load i64, ptr addrspace(1) %"75", align 8
store i64 %"62", ptr addrspace(5) %"58", align 8
%"64" = load i64, ptr addrspace(5) %"56", align 8
%"76" = inttoptr i64 %"64" to ptr addrspace(1)
%"41" = getelementptr inbounds i8, ptr addrspace(1) %"76", i64 8
%"65" = load i64, ptr addrspace(1) %"41", align 4
store i64 %"65", ptr addrspace(5) %"59", align 4
%"67" = load i64, ptr addrspace(5) %"58", align 4
%"68" = load i64, ptr addrspace(5) %"59", align 4
%"65" = load i64, ptr addrspace(1) %"41", align 8
store i64 %"65", ptr addrspace(5) %"59", align 8
%"67" = load i64, ptr addrspace(5) %"58", align 8
%"68" = load i64, ptr addrspace(5) %"59", align 8
%"77" = call i64 @set_shared_temp1(i64 %"67", i64 %"68")
store i64 %"77", ptr addrspace(5) %"59", align 4
store i64 %"77", ptr addrspace(5) %"59", align 8
br label %"46"
"46": ; preds = %"45"
%"69" = load i64, ptr addrspace(5) %"57", align 4
%"70" = load i64, ptr addrspace(5) %"59", align 4
%"69" = load i64, ptr addrspace(5) %"57", align 8
%"70" = load i64, ptr addrspace(5) %"59", align 8
%"79" = inttoptr i64 %"69" to ptr
store i64 %"70", ptr %"79", align 4
store i64 %"70", ptr %"79", align 8
ret void
}

28
ptx/src/test/ll/shared_variable.ll
View File

@ -11,22 +11,22 @@ define amdgpu_kernel void @shared_variable(ptr addrspace(4) byref(i64) %"32", pt
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr addrspace(1)
%"40" = load i64, ptr addrspace(1) %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"42" = load i64, ptr addrspace(5) %"36", align 4
store i64 %"42", ptr addrspace(3) @shared_mem1, align 4
%"43" = load i64, ptr addrspace(3) @shared_mem1, align 4
store i64 %"43", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
%"40" = load i64, ptr addrspace(1) %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"42" = load i64, ptr addrspace(5) %"36", align 8
store i64 %"42", ptr addrspace(3) @shared_mem1, align 8
%"43" = load i64, ptr addrspace(3) @shared_mem1, align 8
store i64 %"43", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"49" = inttoptr i64 %"44" to ptr addrspace(1)
store i64 %"45", ptr addrspace(1) %"49", align 4
store i64 %"45", ptr addrspace(1) %"49", align 8
ret void
}

18
ptx/src/test/ll/shfl_sync_bfly_b32_pred.ll
View File

@ -14,8 +14,8 @@ define amdgpu_kernel void @shfl_sync_bfly_b32_pred(ptr addrspace(4) byref(i64) %
br label %"39"
"39": ; preds = %1
%"48" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"48", ptr addrspace(5) %"43", align 4
%"48" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"48", ptr addrspace(5) %"43", align 8
%"33" = call i32 @__zluda_ptx_impl_sreg_tid(i8 0)
br label %"40"
@ -40,15 +40,15 @@ define amdgpu_kernel void @shfl_sync_bfly_b32_pred(ptr addrspace(4) byref(i64) %
"15": ; preds = %"14", %"40"
%"57" = load i32, ptr addrspace(5) %"45", align 4
%"56" = zext i32 %"57" to i64
store i64 %"56", ptr addrspace(5) %"44", align 4
%"59" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"56", ptr addrspace(5) %"44", align 8
%"59" = load i64, ptr addrspace(5) %"44", align 8
%"58" = mul i64 %"59", 4
store i64 %"58", ptr addrspace(5) %"44", align 4
%"61" = load i64, ptr addrspace(5) %"43", align 4
%"62" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"58", ptr addrspace(5) %"44", align 8
%"61" = load i64, ptr addrspace(5) %"43", align 8
%"62" = load i64, ptr addrspace(5) %"44", align 8
%"60" = add i64 %"61", %"62"
store i64 %"60", ptr addrspace(5) %"43", align 4
%"63" = load i64, ptr addrspace(5) %"43", align 4
store i64 %"60", ptr addrspace(5) %"43", align 8
%"63" = load i64, ptr addrspace(5) %"43", align 8
%"64" = load i32, ptr addrspace(5) %"46", align 4
%"67" = inttoptr i64 %"63" to ptr
store i32 %"64", ptr %"67", align 4

18
ptx/src/test/ll/shfl_sync_down_b32_pred.ll
View File

@ -14,8 +14,8 @@ define amdgpu_kernel void @shfl_sync_down_b32_pred(ptr addrspace(4) byref(i64) %
br label %"39"
"39": ; preds = %1
%"48" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"48", ptr addrspace(5) %"43", align 4
%"48" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"48", ptr addrspace(5) %"43", align 8
%"33" = call i32 @__zluda_ptx_impl_sreg_tid(i8 0)
br label %"40"
@ -40,15 +40,15 @@ define amdgpu_kernel void @shfl_sync_down_b32_pred(ptr addrspace(4) byref(i64) %
"15": ; preds = %"14", %"40"
%"57" = load i32, ptr addrspace(5) %"45", align 4
%"56" = zext i32 %"57" to i64
store i64 %"56", ptr addrspace(5) %"44", align 4
%"59" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"56", ptr addrspace(5) %"44", align 8
%"59" = load i64, ptr addrspace(5) %"44", align 8
%"58" = mul i64 %"59", 4
store i64 %"58", ptr addrspace(5) %"44", align 4
%"61" = load i64, ptr addrspace(5) %"43", align 4
%"62" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"58", ptr addrspace(5) %"44", align 8
%"61" = load i64, ptr addrspace(5) %"43", align 8
%"62" = load i64, ptr addrspace(5) %"44", align 8
%"60" = add i64 %"61", %"62"
store i64 %"60", ptr addrspace(5) %"43", align 4
%"63" = load i64, ptr addrspace(5) %"43", align 4
store i64 %"60", ptr addrspace(5) %"43", align 8
%"63" = load i64, ptr addrspace(5) %"43", align 8
%"64" = load i32, ptr addrspace(5) %"46", align 4
%"67" = inttoptr i64 %"63" to ptr
store i32 %"64", ptr %"67", align 4

18
ptx/src/test/ll/shfl_sync_idx_b32_pred.ll
View File

@ -14,8 +14,8 @@ define amdgpu_kernel void @shfl_sync_idx_b32_pred(ptr addrspace(4) byref(i64) %"
br label %"39"
"39": ; preds = %1
%"48" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"48", ptr addrspace(5) %"43", align 4
%"48" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"48", ptr addrspace(5) %"43", align 8
%"33" = call i32 @__zluda_ptx_impl_sreg_tid(i8 0)
br label %"40"
@ -40,15 +40,15 @@ define amdgpu_kernel void @shfl_sync_idx_b32_pred(ptr addrspace(4) byref(i64) %"
"15": ; preds = %"14", %"40"
%"57" = load i32, ptr addrspace(5) %"45", align 4
%"56" = zext i32 %"57" to i64
store i64 %"56", ptr addrspace(5) %"44", align 4
%"59" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"56", ptr addrspace(5) %"44", align 8
%"59" = load i64, ptr addrspace(5) %"44", align 8
%"58" = mul i64 %"59", 4
store i64 %"58", ptr addrspace(5) %"44", align 4
%"61" = load i64, ptr addrspace(5) %"43", align 4
%"62" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"58", ptr addrspace(5) %"44", align 8
%"61" = load i64, ptr addrspace(5) %"43", align 8
%"62" = load i64, ptr addrspace(5) %"44", align 8
%"60" = add i64 %"61", %"62"
store i64 %"60", ptr addrspace(5) %"43", align 4
%"63" = load i64, ptr addrspace(5) %"43", align 4
store i64 %"60", ptr addrspace(5) %"43", align 8
%"63" = load i64, ptr addrspace(5) %"43", align 8
%"64" = load i32, ptr addrspace(5) %"46", align 4
%"67" = inttoptr i64 %"63" to ptr
store i32 %"64", ptr %"67", align 4

18
ptx/src/test/ll/shfl_sync_mode_b32.ll
View File

@ -20,8 +20,8 @@ define amdgpu_kernel void @shfl_sync_mode_b32(ptr addrspace(4) byref(i64) %"48")
br label %"45"
"45": ; preds = %1
%"54" = load i64, ptr addrspace(4) %"48", align 4
store i64 %"54", ptr addrspace(5) %"49", align 4
%"54" = load i64, ptr addrspace(4) %"48", align 8
store i64 %"54", ptr addrspace(5) %"49", align 8
%"31" = call i32 @__zluda_ptx_impl_sreg_tid(i8 0)
br label %"46"
@ -55,15 +55,15 @@ define amdgpu_kernel void @shfl_sync_mode_b32(ptr addrspace(4) byref(i64) %"48")
store i32 %"72", ptr addrspace(5) %"53", align 4
%"76" = load i32, ptr addrspace(5) %"51", align 4
%"75" = zext i32 %"76" to i64
store i64 %"75", ptr addrspace(5) %"50", align 4
%"78" = load i64, ptr addrspace(5) %"50", align 4
store i64 %"75", ptr addrspace(5) %"50", align 8
%"78" = load i64, ptr addrspace(5) %"50", align 8
%"77" = mul i64 %"78", 4
store i64 %"77", ptr addrspace(5) %"50", align 4
%"80" = load i64, ptr addrspace(5) %"49", align 4
%"81" = load i64, ptr addrspace(5) %"50", align 4
store i64 %"77", ptr addrspace(5) %"50", align 8
%"80" = load i64, ptr addrspace(5) %"49", align 8
%"81" = load i64, ptr addrspace(5) %"50", align 8
%"79" = add i64 %"80", %"81"
store i64 %"79", ptr addrspace(5) %"49", align 4
%"82" = load i64, ptr addrspace(5) %"49", align 4
store i64 %"79", ptr addrspace(5) %"49", align 8
%"82" = load i64, ptr addrspace(5) %"49", align 8
%"83" = load i32, ptr addrspace(5) %"53", align 4
%"92" = inttoptr i64 %"82" to ptr
store i32 %"83", ptr %"92", align 4

18
ptx/src/test/ll/shfl_sync_up_b32_pred.ll
View File

@ -14,8 +14,8 @@ define amdgpu_kernel void @shfl_sync_up_b32_pred(ptr addrspace(4) byref(i64) %"4
br label %"39"
"39": ; preds = %1
%"48" = load i64, ptr addrspace(4) %"42", align 4
store i64 %"48", ptr addrspace(5) %"43", align 4
%"48" = load i64, ptr addrspace(4) %"42", align 8
store i64 %"48", ptr addrspace(5) %"43", align 8
%"33" = call i32 @__zluda_ptx_impl_sreg_tid(i8 0)
br label %"40"
@ -40,15 +40,15 @@ define amdgpu_kernel void @shfl_sync_up_b32_pred(ptr addrspace(4) byref(i64) %"4
"15": ; preds = %"14", %"40"
%"57" = load i32, ptr addrspace(5) %"45", align 4
%"56" = zext i32 %"57" to i64
store i64 %"56", ptr addrspace(5) %"44", align 4
%"59" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"56", ptr addrspace(5) %"44", align 8
%"59" = load i64, ptr addrspace(5) %"44", align 8
%"58" = mul i64 %"59", 4
store i64 %"58", ptr addrspace(5) %"44", align 4
%"61" = load i64, ptr addrspace(5) %"43", align 4
%"62" = load i64, ptr addrspace(5) %"44", align 4
store i64 %"58", ptr addrspace(5) %"44", align 8
%"61" = load i64, ptr addrspace(5) %"43", align 8
%"62" = load i64, ptr addrspace(5) %"44", align 8
%"60" = add i64 %"61", %"62"
store i64 %"60", ptr addrspace(5) %"43", align 4
%"63" = load i64, ptr addrspace(5) %"43", align 4
store i64 %"60", ptr addrspace(5) %"43", align 8
%"63" = load i64, ptr addrspace(5) %"43", align 8
%"64" = load i32, ptr addrspace(5) %"46", align 4
%"67" = inttoptr i64 %"63" to ptr
store i32 %"64", ptr %"67", align 4

24
ptx/src/test/ll/shl.ll
View File

@ -9,22 +9,22 @@ define amdgpu_kernel void @shl(ptr addrspace(4) byref(i64) %"32", ptr addrspace(
br label %"31"
"31": ; preds = %1
%"38" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"38", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(4) %"33", align 4
store i64 %"39", ptr addrspace(5) %"35", align 4
%"41" = load i64, ptr addrspace(5) %"34", align 4
%"38" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"38", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(4) %"33", align 8
store i64 %"39", ptr addrspace(5) %"35", align 8
%"41" = load i64, ptr addrspace(5) %"34", align 8
%"46" = inttoptr i64 %"41" to ptr
%"40" = load i64, ptr %"46", align 4
store i64 %"40", ptr addrspace(5) %"36", align 4
%"43" = load i64, ptr addrspace(5) %"36", align 4
%"40" = load i64, ptr %"46", align 8
store i64 %"40", ptr addrspace(5) %"36", align 8
%"43" = load i64, ptr addrspace(5) %"36", align 8
%2 = shl i64 %"43", 2
%"47" = select i1 false, i64 0, i64 %2
store i64 %"47", ptr addrspace(5) %"37", align 4
%"44" = load i64, ptr addrspace(5) %"35", align 4
%"45" = load i64, ptr addrspace(5) %"37", align 4
store i64 %"47", ptr addrspace(5) %"37", align 8
%"44" = load i64, ptr addrspace(5) %"35", align 8
%"45" = load i64, ptr addrspace(5) %"37", align 8
%"49" = inttoptr i64 %"44" to ptr
store i64 %"45", ptr %"49", align 4
store i64 %"45", ptr %"49", align 8
ret void
}

12
ptx/src/test/ll/shr.ll
View File

@ -8,11 +8,11 @@ define amdgpu_kernel void @shr(ptr addrspace(4) byref(i64) %"31", ptr addrspace(
br label %"30"
"30": ; preds = %1
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"37" = load i64, ptr addrspace(4) %"32", align 4
store i64 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"37" = load i64, ptr addrspace(4) %"32", align 8
store i64 %"37", ptr addrspace(5) %"34", align 8
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"44" = inttoptr i64 %"39" to ptr
%"38" = load i32, ptr %"44", align 4
store i32 %"38", ptr addrspace(5) %"35", align 4
@ -20,7 +20,7 @@ define amdgpu_kernel void @shr(ptr addrspace(4) byref(i64) %"31", ptr addrspace(
%2 = ashr i32 %"41", 1
%"40" = select i1 false, i32 0, i32 %2
store i32 %"40", ptr addrspace(5) %"35", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 4
%"42" = load i64, ptr addrspace(5) %"34", align 8
%"43" = load i32, ptr addrspace(5) %"35", align 4
%"45" = inttoptr i64 %"42" to ptr
store i32 %"43", ptr %"45", align 4

12
ptx/src/test/ll/sign_extend.ll
View File

@ -8,16 +8,16 @@ define amdgpu_kernel void @sign_extend(ptr addrspace(4) byref(i64) %"30", ptr ad
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"42" = inttoptr i64 %"38" to ptr
%"41" = load i16, ptr %"42", align 2
%"37" = sext i16 %"41" to i32
store i32 %"37", ptr addrspace(5) %"34", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 4
%"39" = load i64, ptr addrspace(5) %"33", align 8
%"40" = load i32, ptr addrspace(5) %"34", align 4
%"43" = inttoptr i64 %"39" to ptr
store i32 %"40", ptr %"43", align 4

12
ptx/src/test/ll/sin.ll
View File

@ -8,18 +8,18 @@ define amdgpu_kernel void @sin(ptr addrspace(4) byref(i64) %"30", ptr addrspace(
br label %"29"
"29": ; preds = %1
%"35" = load i64, ptr addrspace(4) %"30", align 4
store i64 %"35", ptr addrspace(5) %"32", align 4
%"36" = load i64, ptr addrspace(4) %"31", align 4
store i64 %"36", ptr addrspace(5) %"33", align 4
%"38" = load i64, ptr addrspace(5) %"32", align 4
%"35" = load i64, ptr addrspace(4) %"30", align 8
store i64 %"35", ptr addrspace(5) %"32", align 8
%"36" = load i64, ptr addrspace(4) %"31", align 8
store i64 %"36", ptr addrspace(5) %"33", align 8
%"38" = load i64, ptr addrspace(5) %"32", align 8
%"43" = inttoptr i64 %"38" to ptr
%"37" = load float, ptr %"43", align 4
store float %"37", ptr addrspace(5) %"34", align 4
%"40" = load float, ptr addrspace(5) %"34", align 4
%"39" = call afn float @llvm.sin.f32(float %"40")
store float %"39", ptr addrspace(5) %"34", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 4
%"41" = load i64, ptr addrspace(5) %"33", align 8
%"42" = load float, ptr addrspace(5) %"34", align 4
%"44" = inttoptr i64 %"41" to ptr
store float %"42", ptr %"44", align 4

Some files were not shown because too many files have changed in this diff Show More