AMDGPU: keep track of modifiers when converting v_mac to v_mad

[hakzsam]

Summary

Since v_max_f32_e64/v_max_f16_e64 can be folded if the target
instruction supports the clamp bit, we also need to maintain
modifiers when converting v_mac to v_mad.

This fixes a rendering issue with Dirt Rally because a v_mac
instruction with the clamp bit set was converted to a v_mad
but that bit was lost during the conversion.

Fixes: e184e01dd79 ("AMDGPU: Fold FP clamp as modifier bit")

[arsenm]

Needs tests

lib/Target/AMDGPU/SIInstrInfo.cpp

1779

hasModifiersSet won't preserve the value, you need to keep the whole operand. These aren't simple booleans.

1785–1786

should just add the value. omod is not a simple boolean so that is also broken

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

1783

This is also unnecessary, there is no src2_modifiers for mac so this can stay add 0

[hakzsam]

v2: - preserve valye by using getNamedOperand()->getImm() instead

• add v_clamp_mac_to_mad test in clamp-modifier.ll
• add v_omod_mac_to_mad test in omod.ll

[arsenm]

LGTM

[arsenm]

r297556

lib/Target/AMDGPU/SIInstrInfo.cpp

 
   const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
   const MachineOperand *Src0 = getNamedOperand(MI, AMDGPU::OpName::src0);
+  const MachineOperand *Src0Mods =
+    getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
   const MachineOperand *Src1 = getNamedOperand(MI, AMDGPU::OpName::src1);
+  const MachineOperand *Src1Mods =
+    getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
   const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
+  const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
+  const MachineOperand *Omod = getNamedOperand(MI, AMDGPU::OpName::omod);
 
   return BuildMI(*MBB, MI, MI.getDebugLoc(),
                  get(IsF16 ? AMDGPU::V_MAD_F16 : AMDGPU::V_MAD_F32))
       .add(*Dst)
-      .addImm(0) // Src0 mods
+      .addImm(Src0Mods ? Src0Mods->getImm() : 0)
       .add(*Src0)
-      .addImm(0) // Src1 mods
+      .addImm(Src1Mods ? Src1Mods->getImm() : 0)
       .add(*Src1)
       .addImm(0) // Src mods
       .add(*Src2)
-      .addImm(0)  // clamp
-      .addImm(0); // omod
+      .addImm(Clamp ? Clamp->getImm() : 0)
+      .addImm(Omod ? Omod->getImm() : 0);
 }
 
 // It's not generally safe to move VALU instructions across these since it will

test/CodeGen/AMDGPU/clamp-modifier.ll

   ret void
 }
 
+; GCN-LABEL: {{^}}v_clamp_mac_to_mad:
+; GCN: v_mad_f32 v{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]}} clamp{{$}}
+define amdgpu_kernel void @v_clamp_mac_to_mad(float addrspace(1)* %out, float addrspace(1)* %aptr, float %a) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %b = load float, float addrspace(1)* %gep0
+
+  %mul = fmul float %a, %a
+  %add = fadd float %mul, %b
+  %max = call float @llvm.maxnum.f32(float %add, float 0.0)
+  %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
+  %res = fadd float %clamp, %b
+  store float %res, float addrspace(1)* %out.gep
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare float @llvm.fabs.f32(float) #1
 declare float @llvm.floor.f32(float) #1

test/CodeGen/AMDGPU/omod.ll

   ret void
 }
 
+; GCN-LABEL: {{^}}v_omod_mac_to_mad:
+; GCN: v_mad_f32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]}} mul:2{{$}}
+define amdgpu_ps void @v_omod_mac_to_mad(float %b, float %a) #0 {
+  %mul = fmul float %a, %a
+  %add = fadd float %mul, %b
+  %mad = fmul float %add, 2.0
+  %res = fmul float %mad, %b
+  store float %res, float addrspace(1)* undef
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare float @llvm.fabs.f32(float) #1
 declare float @llvm.floor.f32(float) #1