AMDGPU: Ignore any_extend in mul24 combine

[arsenm]

Summary

If a multiply is truncated, SimplifyDemandedBits
sometimes turns a zero_extend of the inputs into an
any_extend, which makes the known bits computation unhelpful.
Ignore these and compute known bits for the underlying value,
since we insert the correct extend type after.

[kzhuravl]

LGTM

[rampitec]

LGTM

[arsenm]

r331919

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
   if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16))
     return SDValue();
 
   SelectionDAG &DAG = DCI.DAG;
   SDLoc DL(N);
 
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
+
+  // SimplifyDemandedBits has the annoying habit of turning useful zero_extends
+  // in the source into any_extends if the result of the mul is truncated. Since
+  // we can assume the high bits are whatever we want, use the underlying value
+  // to avoid the unknown high bits from interfering.
+  if (N0.getOpcode() == ISD::ANY_EXTEND)
+    N0 = N0.getOperand(0);
+
+  if (N1.getOpcode() == ISD::ANY_EXTEND)
+    N1 = N1.getOperand(0);
+
   SDValue Mul;
 
   if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
     N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
     N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
     Mul = getMul24(DAG, DL, N0, N1, Size, false);
   } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
     N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);

test/CodeGen/AMDGPU/mul_uint24-amdgcn.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=FUNC %s
-; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI,FUNC %s
 
 declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
 declare i32 @llvm.amdgcn.workitem.id.y() nounwind readnone
 
 ; FUNC-LABEL: {{^}}test_umul24_i32:
 ; GCN: v_mul_u32_u24
 define amdgpu_kernel void @test_umul24_i32(i32 addrspace(1)* %out, i32 %a, i32 %b) {
 entry:
   %0 = shl i32 %a, 8
   %a_24 = lshr i32 %0, 8
   %1 = shl i32 %b, 8
   %b_24 = lshr i32 %1, 8
   %2 = mul i32 %a_24, %b_24
   store i32 %2, i32 addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}test_umul24_i16_sext:
-; SI: v_mul_u32_u24_e{{(32|64)}} [[VI_MUL:v[0-9]]], {{[sv][0-9], [sv][0-9]}}
-; SI: v_bfe_i32 v{{[0-9]}}, [[VI_MUL]], 0, 16
-; VI: s_mul_i32 [[SI_MUL:s[0-9]]], s{{[0-9]}}, s{{[0-9]}}
-; VI: s_sext_i32_i16 s{{[0-9]}}, [[SI_MUL]]
+; GCN: v_mul_u32_u24_e{{(32|64)}} [[VI_MUL:v[0-9]]], {{[sv][0-9], [sv][0-9]}}
+; GCN: v_bfe_i32 v{{[0-9]}}, [[VI_MUL]], 0, 16
 define amdgpu_kernel void @test_umul24_i16_sext(i32 addrspace(1)* %out, i16 %a, i16 %b) {
 entry:
   %mul = mul i16 %a, %b
   %ext = sext i16 %mul to i32
   store i32 %ext, i32 addrspace(1)* %out
   ret void
 }
 
define amdgpu_kernel void @test_umul24_i16_vgpr_sext(i32 addrspace(1)* %out, i16 addrspace(1)* %in) {
   %b = load i16, i16 addrspace(1)* %ptr_b
   %mul = mul i16 %a, %b
   %val = sext i16 %mul to i32
   store i32 %val, i32 addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}test_umul24_i16:
-; SI: s_and_b32
-; SI: v_mul_u32_u24_e32
-; SI: v_and_b32_e32
-; VI: s_mul_i32
-; VI: s_and_b32
-; VI: v_mov_b32_e32
+; GCN: s_and_b32
+; GCN: v_mul_u32_u24_e32
+; GCN: v_and_b32_e32
 define amdgpu_kernel void @test_umul24_i16(i32 addrspace(1)* %out, i16 %a, i16 %b) {
 entry:
   %mul = mul i16 %a, %b
   %ext = zext i16 %mul to i32
   store i32 %ext, i32 addrspace(1)* %out
   ret void
 }
 
entry:
   %tmp1 = shl i33 %b, 9
   %b_24 = lshr i33 %tmp1, 9
   %tmp2 = mul i33 %a_24, %b_24
   %hi = lshr i33 %tmp2, 32
   %trunc = trunc i33 %hi to i32
   store i32 %trunc, i32 addrspace(1)* %out
   ret void
 }
+
+
+; Make sure the created any_extend is ignored to use the real bits
+; being multiplied.
+
+; GCN-LABEL: {{^}}test_umul24_anyextend_i24_src0_src1:
+; GCN-DAG: v_mul_u32_u24_e32 v0, 0xea, v0
+; GCN-DAG: v_mul_u32_u24_e32 v1, 0x39b, v1
+; GCN: v_mul_u32_u24_e32 v0, v0, v1
+; GCN: v_and_b32_e32 v0, 0x1fffe, v0
+; GCN: v_mul_u32_u24_e32 v0, 0x63, v0
+; GCN: s_setpc_b64
+define i17 @test_umul24_anyextend_i24_src0_src1(i24 %a, i24 %b) {
+entry:
+  %aa = mul i24 %a, 234
+  %bb = mul i24 %b, 923
+  %a_32 = zext i24 %aa to i32
+  %b_32 = zext i24 %bb to i32
+  %mul = mul i32 %a_32, %b_32
+  %trunc = trunc i32 %mul to i17
+  %arst = mul i17 %trunc, 99
+  ret i17 %arst
+}
+
+; GCN-LABEL: {{^}}test_umul24_anyextend_i23_src0_src1:
+; GCN: s_mov_b32 [[U23_MASK:s[0-9]+]], 0x7fffff
+; GCN-DAG: v_and_b32_e32 v0, [[U23_MASK]], v0
+; GCN-DAG: v_and_b32_e32 v1, [[U23_MASK]], v1
+; GCN-DAG: v_mul_u32_u24_e32 v0, 0xea, v0
+; GCN-DAG: v_mul_u32_u24_e32 v1, 0x39b, v1
+; GCN: v_and_b32_e32 v1, s6, v1
+; GCN: v_and_b32_e32 v0, 0x7ffffe, v0
+; GCN: v_mul_u32_u24_e32 v0, v0, v1
+; GCN: v_and_b32_e32 v0, 0x1fffe, v0
+; GCN: v_mul_u32_u24_e32 v0, 0x63, v0
+; GCN: s_setpc_b64
+define i17 @test_umul24_anyextend_i23_src0_src1(i23 %a, i23 %b) {
+entry:
+  %aa = mul i23 %a, 234
+  %bb = mul i23 %b, 923
+  %a_32 = zext i23 %aa to i32
+  %b_32 = zext i23 %bb to i32
+  %mul = mul i32 %a_32, %b_32
+  %trunc = trunc i32 %mul to i17
+  %arst = mul i17 %trunc, 99
+  ret i17 %arst
+}