[InstCombine][AMDGPU] Fix crash with v3i16/v3f16 buffer intrinsics

[piotr]

Summary

This is something of a workaround to avoid a crash later on in type
legalizer (WidenVectorResult()).
Also added some f16 tests, including a non-working v3f16 case with
a FIXME.

[piotr]

Fixed wrong spacing.

[arsenm]

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

975

I think I have these working in GlobalISel already

977

!= 32 would be a bit safer

[piotr]

Changed "== 16" to "!= 32".

[arsenm]

LGTM

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

 
 /// Implement SimplifyDemandedVectorElts for amdgcn buffer and image intrinsics.
 ///
 /// Note: This only supports non-TFE/LWE image intrinsic calls; those have
 ///       struct returns.
 Value *InstCombiner::simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II,
                                                            APInt DemandedElts,
                                                            int DMaskIdx) {
+
+  // FIXME: Allow v3i16/v3f16 in buffer intrinsics when the types are fully supported.
+  if (DMaskIdx < 0 &&
+      II->getType()->getScalarSizeInBits() == 16 &&
+      DemandedElts.getActiveBits() == 3)
+    return nullptr;
+
   unsigned VWidth = II->getType()->getVectorNumElements();
   if (VWidth == 1)
     return nullptr;
 
   ConstantInt *NewDMask = nullptr;
 
   if (DMaskIdx < 0) {
     // Pretend that a prefix of elements is demanded to simplify the code

test/Transforms/InstCombine/AMDGPU/amdgcn-demanded-vector-elts.ll

 
 declare float @llvm.amdgcn.raw.tbuffer.load.f32(<4 x i32>, i32, i32, i32, i32) #1
 declare <2 x float> @llvm.amdgcn.raw.tbuffer.load.v2f32(<4 x i32>, i32, i32, i32, i32) #1
 declare <3 x float> @llvm.amdgcn.raw.tbuffer.load.v3f32(<4 x i32>, i32, i32, i32, i32) #1
 declare <4 x float> @llvm.amdgcn.raw.tbuffer.load.v4f32(<4 x i32>, i32, i32, i32, i32) #1
 
 declare <4 x i32> @llvm.amdgcn.raw.tbuffer.load.v4i32(<4 x i32>, i32, i32, i32, i32) #1
 
+; CHECK-LABEL: @extract_elt3_raw_tbuffer_load_v4f16(
+; CHECK-NEXT: %data = call <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+; CHECK-NEXT: %elt1 = extractelement <4 x half> %data, i32 3
+; CHECK-NEXT: ret half %elt1
+define amdgpu_ps half @extract_elt3_raw_tbuffer_load_v4f16(<4 x i32> inreg %rsrc, i32 %arg0, i32 inreg %arg1) #0 {
+  %data = call <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+  %elt1 = extractelement <4 x half> %data, i32 3
+  ret half %elt1
+}
+
+; FIXME: Enable load shortening when full support for v3f16 has been added (should expect call <3 x half> @llvm.amdgcn.raw.tbuffer.load.v3f16).
+; CHECK-LABEL: @extract_elt2_raw_tbuffer_load_v4f16(
+; CHECK-NEXT: %data = call <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+; CHECK-NEXT: %elt1 = extractelement <4 x half> %data, i32 2
+; CHECK-NEXT: ret half %elt1
+define amdgpu_ps half @extract_elt2_raw_tbuffer_load_v4f16(<4 x i32> inreg %rsrc, i32 %arg0, i32 inreg %arg1) #0 {
+  %data = call <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+  %elt1 = extractelement <4 x half> %data, i32 2
+  ret half %elt1
+}
+
+; CHECK-LABEL: @extract_elt1_raw_tbuffer_load_v4f16(
+; CHECK-NEXT: %data = call <2 x half> @llvm.amdgcn.raw.tbuffer.load.v2f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+; CHECK-NEXT: %elt1 = extractelement <2 x half> %data, i32 1
+; CHECK-NEXT: ret half %elt1
+define amdgpu_ps half @extract_elt1_raw_tbuffer_load_v4f16(<4 x i32> inreg %rsrc, i32 %arg0, i32 inreg %arg1) #0 {
+  %data = call <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+  %elt1 = extractelement <4 x half> %data, i32 1
+  ret half %elt1
+}
+
+; CHECK-LABEL: @extract_elt0_raw_tbuffer_load_v4f16(
+; CHECK-NEXT: %data = call half @llvm.amdgcn.raw.tbuffer.load.f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+; CHECK-NEXT: ret half %data
+define amdgpu_ps half @extract_elt0_raw_tbuffer_load_v4f16(<4 x i32> inreg %rsrc, i32 %arg0, i32 inreg %arg1) #0 {
+  %data = call <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 78, i32 0)
+  %elt1 = extractelement <4 x half> %data, i32 0
+  ret half %elt1
+}
+
+declare half @llvm.amdgcn.raw.tbuffer.load.f16(<4 x i32>, i32, i32, i32, i32) #1
+declare <2 x half> @llvm.amdgcn.raw.tbuffer.load.v2f16(<4 x i32>, i32, i32, i32, i32) #1
+declare <3 x half> @llvm.amdgcn.raw.tbuffer.load.v3f16(<4 x i32>, i32, i32, i32, i32) #1
+declare <4 x half> @llvm.amdgcn.raw.tbuffer.load.v4f16(<4 x i32>, i32, i32, i32, i32) #1
+
 ; --------------------------------------------------------------------
 ; llvm.amdgcn.struct.tbuffer.load
 ; --------------------------------------------------------------------
 
 ; CHECK-LABEL: @struct_tbuffer_load_f32(
 ; CHECK-NEXT: %data = call float @llvm.amdgcn.struct.tbuffer.load.f32(<4 x i32> %rsrc, i32 %arg0, i32 %arg1, i32 %arg2, i32 78, i32 0)
 ; CHECK-NEXT: ret float %data
 define amdgpu_ps float @struct_tbuffer_load_f32(<4 x i32> inreg %rsrc, i32 %arg0, i32 %arg1, i32 inreg %arg2) #0 {