AMDGPU: Fold shift into cvt_f32_ubyteN

[• tstellarAMD]

lib/Target/AMDGPU/SIISelLowering.cpp

2678–2681

The comment doesn't match the if statement. Which is correct?

[arsenm]

lib/Target/AMDGPU/SIISelLowering.cpp

2678–2681

Should should be SRLs

[• tstellarAMD]

With the comment fix, LGTM.

[arsenm]

r268930

lib/Target/AMDGPU/SIISelLowering.cpp

case AMDGPUISD::FMAX_LEGACY: {
     break;
   }
 
   case AMDGPUISD::CVT_F32_UBYTE0:
   case AMDGPUISD::CVT_F32_UBYTE1:
   case AMDGPUISD::CVT_F32_UBYTE2:
   case AMDGPUISD::CVT_F32_UBYTE3: {
     unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
-
     SDValue Src = N->getOperand(0);
+
+    if (Src.getOpcode() == ISD::SRL) {
+      // cvt_f32_ubyte0 (shl x, 16) -> cvt_f32_ubyte2 x
+      // cvt_f32_ubyte1 (shl x, 16) -> cvt_f32_ubyte3 x
+      // cvt_f32_ubyte0 (shl x, 8) -> cvt_f32_ubyte1 x
+
+      if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src.getOperand(1))) {
+        unsigned SrcOffset = C->getZExtValue() + 8 * Offset;
+        if (SrcOffset < 32 && SrcOffset % 8 == 0) {
+          return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0 + SrcOffset / 8, DL,
+                             MVT::f32, Src.getOperand(0));
+        }
+      }
+    }
+
     APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
 
     APInt KnownZero, KnownOne;
     TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
                                           !DCI.isBeforeLegalizeOps());
     const TargetLowering &TLI = DAG.getTargetLoweringInfo();
     if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
         TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {

test/CodeGen/AMDGPU/cvt_f32_ubyte.ll

-; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 
 ; SI-LABEL: {{^}}load_i8_to_f32:
 ; SI: buffer_load_ubyte [[LOADREG:v[0-9]+]],
 ; SI-NOT: bfe
 ; SI-NOT: lshr
 ; SI: v_cvt_f32_ubyte0_e32 [[CONV:v[0-9]+]], [[LOADREG]]
 ; SI: buffer_store_dword [[CONV]],
define void @i8_zext_inreg_hi1_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
   %load = load i32, i32 addrspace(1)* %in, align 4
   %inreg = and i32 %load, 65280
   %shr = lshr i32 %inreg, 8
   %cvt = uitofp i32 %shr to float
   store float %cvt, float addrspace(1)* %out, align 4
   ret void
 }
 
-
 ; We don't get these ones because of the zext, but instcombine removes
 ; them so it shouldn't really matter.
+; SI-LABEL: {{^}}i8_zext_i32_to_f32:
 define void @i8_zext_i32_to_f32(float addrspace(1)* noalias %out, i8 addrspace(1)* noalias %in) nounwind {
   %load = load i8, i8 addrspace(1)* %in, align 1
   %ext = zext i8 %load to i32
   %cvt = uitofp i32 %ext to float
   store float %cvt, float addrspace(1)* %out, align 4
   ret void
 }
 
+; SI-LABEL: {{^}}v4i8_zext_v4i32_to_v4f32:
 define void @v4i8_zext_v4i32_to_v4f32(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %in) nounwind {
   %load = load <4 x i8>, <4 x i8> addrspace(1)* %in, align 1
   %ext = zext <4 x i8> %load to <4 x i32>
   %cvt = uitofp <4 x i32> %ext to <4 x float>
   store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
   ret void
 }
+
+; SI-LABEL: {{^}}extract_byte0_to_f32:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte0_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @extract_byte0_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in
+  %and = and i32 %val, 255
+  %cvt = uitofp i32 %and to float
+  store float %cvt, float addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}extract_byte1_to_f32:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte1_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @extract_byte1_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in
+  %srl = lshr i32 %val, 8
+  %and = and i32 %srl, 255
+  %cvt = uitofp i32 %and to float
+  store float %cvt, float addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}extract_byte2_to_f32:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte2_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @extract_byte2_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in
+  %srl = lshr i32 %val, 16
+  %and = and i32 %srl, 255
+  %cvt = uitofp i32 %and to float
+  store float %cvt, float addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}extract_byte3_to_f32:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte3_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @extract_byte3_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in
+  %srl = lshr i32 %val, 24
+  %and = and i32 %srl, 255
+  %cvt = uitofp i32 %and to float
+  store float %cvt, float addrspace(1)* %out
+  ret void
+}

test/CodeGen/AMDGPU/llvm.AMDGPU.cvt_f32_ubyte.ll

 ; SI-LABEL: {{^}}test_unpack_byte3_to_float:
 ; SI: v_cvt_f32_ubyte3
 define void @test_unpack_byte3_to_float(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
   %val = load i32, i32 addrspace(1)* %in, align 4
   %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte3(i32 %val) nounwind readnone
   store float %cvt, float addrspace(1)* %out, align 4
   ret void
 }
+
+; SI-LABEL: {{^}}byte1_shift8:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte2_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @byte1_shift8(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %shift = lshr i32 %val, 8
+  %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte1(i32 %shift) nounwind readnone
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}byte1_shift7:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI: v_lshrrev_b32_e32 [[SRL:v[0-9]+]], 7, [[VAL]]
+; SI: v_cvt_f32_ubyte1_e32 [[CONV:v[0-9]+]], [[SRL]]
+; SI: buffer_store_dword [[CONV]]
+define void @byte1_shift7(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %shift = lshr i32 %val, 7
+  %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte1(i32 %shift) nounwind readnone
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}byte1_shift16:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte3_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @byte1_shift16(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %shift = lshr i32 %val, 16
+  %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte1(i32 %shift) nounwind readnone
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}byte2_shift8:
+; SI: buffer_load_dword [[VAL:v[0-9]+]]
+; SI-NOT: [[VAL]]
+; SI: v_cvt_f32_ubyte3_e32 [[CONV:v[0-9]+]], [[VAL]]
+; SI: buffer_store_dword [[CONV]]
+define void @byte2_shift8(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %shift = lshr i32 %val, 8
+  %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte2(i32 %shift) nounwind readnone
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; XXX - undef
+; SI-LABEL: {{^}}byte1_shift24:
+; SI: v_cvt_f32_ubyte1_e32 [[CONV:v[0-9]+]], 0
+; SI: buffer_store_dword [[CONV]]
+define void @byte1_shift24(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %shift = lshr i32 %val, 24
+  %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte1(i32 %shift) nounwind readnone
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}