[AMDGPU] Fix ieee mode default value

[Flakebi]

Summary

Previously, the default value for ieee mode was

• on for compute kernels and compute shaders,
• off for all shaders except compute shaders.

This commit changes the default to be

• on for compute kernels,
• off for shaders.

This aligns the default value with the settings that are actually in
use. To my knowledge, all users of shader calling conventions (mesa and
llpc) disable the ieee mode by default.

[foad]

Makes sense to me. As I understand it the biggest practical effect is that the compiler will start making use of output modifiers in compute shaders.

[nhaehnle]

Please make sure you give this a reasonable amount of testing, but it totally makes sense to me.

[arsenm]

Can you add a test that directly checks the mode setting in the output

[Flakebi]

Add pre-commited tests.

@nhaehnle I ran piglit run gpu results/gpu and Vulkan CTS on D89388 and D89399 and didn’t notice any new failures (if these are the right tests?).

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

SIModeRegisterDefaults() :
     FP32InputDenormals(true),
     FP32OutputDenormals(true),
     FP64FP16InputDenormals(true),
     FP64FP16OutputDenormals(true) {}
 
   SIModeRegisterDefaults(const Function &F);
 
   static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) {
-    const bool IsCompute = AMDGPU::isCompute(CC);
-
     SIModeRegisterDefaults Mode;
-    Mode.IEEE = IsCompute;
+    Mode.IEEE = !AMDGPU::isShader(CC);
     return Mode;
   }
 
   bool operator ==(const SIModeRegisterDefaults Other) const {
     return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp &&
            FP32InputDenormals == Other.FP32InputDenormals &&
            FP32OutputDenormals == Other.FP32OutputDenormals &&
            FP64FP16InputDenormals == Other.FP64FP16InputDenormals &&

llvm/test/CodeGen/AMDGPU/amdgcn-ieee.ll

define void @func_ieee_mode_off() #2 {
   %min = call float @llvm.minnum.f32(float %val0, float %val1)
   store volatile float %min, float addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}cs_ieee_mode_default:
 ; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
 ; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
-; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
-; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
+; GCN-NOT: [[VAL0]]
+; GCN-NOT: [[VAL1]]
 ; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
 ; GCN-NOT: v_mul_f32
 define amdgpu_cs void @cs_ieee_mode_default() #0 {
   %val0 = load volatile float, float addrspace(1)* undef
   %val1 = load volatile float, float addrspace(1)* undef
   %min = call float @llvm.minnum.f32(float %val0, float %val1)
   store volatile float %min, float addrspace(1)* undef
   ret void