[AMDGPU] Limit GFX11 to using 128 VGPRs

This is a temporary measure to avoid generating incorrect code until the
compiler understands the new way that GFX11 encodes 16-bit operands in
VOP instructions.

Differential Revision: https://reviews.llvm.org/D128054

Author: jayfoad

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

@@ -33,6 +33,11 @@ static llvm::cl::opt<unsigned>
                             llvm::cl::desc("AMDHSA Code Object Version"),
                             llvm::cl::init(4));
 
+// TODO-GFX11: Remove this when full 16-bit codegen is implemented.
+static llvm::cl::opt<bool>
+    LimitTo128VGPRs("amdgpu-limit-to-128-vgprs", llvm::cl::Hidden,
+                    llvm::cl::desc("Never use more than 128 VGPRs"));
+
 namespace {
 
 /// \returns Bit mask for given bit \p Shift and bit \p Width.
@@ -795,6 +800,15 @@ unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI) {
 }
 
 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI) {
+  if (LimitTo128VGPRs.getNumOccurrences() ? LimitTo128VGPRs
+                                          : isGFX11Plus(*STI)) {
+    // GFX11 changes the encoding of 16-bit operands in VOP1/2/C instructions
+    // such that values 128..255 no longer mean v128..v255, they mean
+    // v0.hi..v127.hi instead. Until the compiler understands this, it is not
+    // safe to use v128..v255.
+    // TODO-GFX11: Remove this when full 16-bit codegen is implemented.
+    return 128;
+  }
   if (STI->getFeatureBits().test(FeatureGFX90AInsts))
     return 512;
   return 256;

llvm/test/CodeGen/AMDGPU/attr-amdgpu-flat-work-group-size-vgpr-limit.ll

@@ -536,10 +536,10 @@ define internal void @use256vgprs() {
 ; GFX10WGP-WAVE64: NumVgprs: 256
 ; GFX10CU-WAVE32: NumVgprs: 256
 ; GFX10CU-WAVE64: NumVgprs: 256
-; GFX11WGP-WAVE32: NumVgprs: 256
-; GFX11WGP-WAVE64: NumVgprs: 256
-; GFX11CU-WAVE32: NumVgprs: 256
-; GFX11CU-WAVE64: NumVgprs: 256
+; GFX11WGP-WAVE32: NumVgprs: 128
+; GFX11WGP-WAVE64: NumVgprs: 128
+; GFX11CU-WAVE32: NumVgprs: 128
+; GFX11CU-WAVE64: NumVgprs: 128
 define amdgpu_kernel void @f256() #256 {
   call void @use256vgprs()
   ret void
@@ -555,8 +555,8 @@ attributes #256 = { nounwind "amdgpu-flat-work-group-size"="256,256" }
 ; GFX10WGP-WAVE64: NumVgprs: 256
 ; GFX10CU-WAVE32: NumVgprs: 128
 ; GFX10CU-WAVE64: NumVgprs: 128
-; GFX11WGP-WAVE32: NumVgprs: 256
-; GFX11WGP-WAVE64: NumVgprs: 256
+; GFX11WGP-WAVE32: NumVgprs: 128
+; GFX11WGP-WAVE64: NumVgprs: 128
 ; GFX11CU-WAVE32: NumVgprs: 128
 ; GFX11CU-WAVE64: NumVgprs: 128
 define amdgpu_kernel void @f512() #512 {