[SPIRV] Add PreLegalizer pattern matching for faceforward
#139959
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
llvmbot
added
clang
Member
|
@llvm/pr-subscribers-llvm-globalisel @llvm/pr-subscribers-backend-spir-v Author: Kaitlin Peng (kmpeng) ChangesTasks completed:
Closes #137255. Patch is 25.07 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/139959.diff 7 Files Affected:
diff --git a/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h b/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h
index 4eb7b8f45c85a..e1996fd3e30ed 100644
--- a/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h
+++ b/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h
@@ -127,11 +127,7 @@ template <typename T> constexpr vector<T, 4> lit_impl(T NDotL, T NDotH, T M) {
}
template <typename T> constexpr T faceforward_impl(T N, T I, T Ng) {
-#if (__has_builtin(__builtin_spirv_faceforward))
- return __builtin_spirv_faceforward(N, I, Ng);
-#else
return select(dot(I, Ng) < 0, N, -N);
-#endif
}
template <typename T> constexpr T ldexp_impl(T X, T Exp) {
diff --git a/clang/test/CodeGenHLSL/builtins/faceforward.hlsl b/clang/test/CodeGenHLSL/builtins/faceforward.hlsl
index d2ece57aba4ae..f12f75f384964 100644
--- a/clang/test/CodeGenHLSL/builtins/faceforward.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/faceforward.hlsl
@@ -12,8 +12,11 @@
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, half %{{.*}}, half %fneg.i
// CHECK: ret half %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef half @llvm.spv.faceforward.f16(half %{{.*}}, half %{{.*}}, half %{{.*}})
-// SPVCHECK: ret half %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = fmul reassoc nnan ninf nsz arcp afn half %{{.*}}, %{{.*}}
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn half %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, half %{{.*}}, half %fneg.i
+// SPVCHECK: ret half %hlsl.select.i
half test_faceforward_half(half N, half I, half Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_half2
@@ -23,8 +26,11 @@ half test_faceforward_half(half N, half I, half Ng) { return faceforward(N, I, N
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x half> %{{.*}}, <2 x half> %fneg.i
// CHECK: ret <2 x half> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half2
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <2 x half> @llvm.spv.faceforward.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, <2 x half> %{{.*}})
-// SPVCHECK: ret <2 x half> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <2 x half> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x half> %{{.*}}, <2 x half> %fneg.i
+// SPVCHECK: ret <2 x half> %hlsl.select.i
half2 test_faceforward_half2(half2 N, half2 I, half2 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_half3
@@ -34,8 +40,11 @@ half2 test_faceforward_half2(half2 N, half2 I, half2 Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x half> %{{.*}}, <3 x half> %fneg.i
// CHECK: ret <3 x half> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half3
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <3 x half> @llvm.spv.faceforward.v3f16(<3 x half> %{{.*}}, <3 x half> %{{.*}}, <3 x half> %{{.*}})
-// SPVCHECK: ret <3 x half> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v3f16(<3 x half> %{{.*}}, <3 x half> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <3 x half> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x half> %{{.*}}, <3 x half> %fneg.i
+// SPVCHECK: ret <3 x half> %hlsl.select.i
half3 test_faceforward_half3(half3 N, half3 I, half3 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_half4
@@ -45,8 +54,11 @@ half3 test_faceforward_half3(half3 N, half3 I, half3 Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x half> %{{.*}}, <4 x half> %fneg.i
// CHECK: ret <4 x half> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half4
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <4 x half> @llvm.spv.faceforward.v4f16(<4 x half> %{{.*}}, <4 x half> %{{.*}}, <4 x half> %{{.*}})
-// SPVCHECK: ret <4 x half> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v4f16(<4 x half> %{{.*}}, <4 x half> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <4 x half> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x half> %{{.*}}, <4 x half> %fneg.i
+// SPVCHECK: ret <4 x half> %hlsl.select.i
half4 test_faceforward_half4(half4 N, half4 I, half4 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float
@@ -56,8 +68,11 @@ half4 test_faceforward_half4(half4 N, half4 I, half4 Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, float %{{.*}}, float %fneg.i
// CHECK: ret float %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef float @llvm.spv.faceforward.f32(float %{{.*}}, float %{{.*}}, float %{{.*}})
-// SPVCHECK: ret float %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = fmul reassoc nnan ninf nsz arcp afn float %{{.*}}, %{{.*}}
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn float %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, float %{{.*}}, float %fneg.i
+// SPVCHECK: ret float %hlsl.select.i
float test_faceforward_float(float N, float I, float Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float2
@@ -67,8 +82,11 @@ float test_faceforward_float(float N, float I, float Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x float> %{{.*}}, <2 x float> %fneg.i
// CHECK: ret <2 x float> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float2
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <2 x float> @llvm.spv.faceforward.v2f32(<2 x float> %{{.*}}, <2 x float> %{{.*}}, <2 x float> %{{.*}})
-// SPVCHECK: ret <2 x float> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fdot.v2f32(<2 x float> %{{.*}}, <2 x float> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <2 x float> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x float> %{{.*}}, <2 x float> %fneg.i
+// SPVCHECK: ret <2 x float> %hlsl.select.i
float2 test_faceforward_float2(float2 N, float2 I, float2 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float3
@@ -78,8 +96,11 @@ float2 test_faceforward_float2(float2 N, float2 I, float2 Ng) { return faceforwa
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x float> %{{.*}}, <3 x float> %fneg.i
// CHECK: ret <3 x float> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float3
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <3 x float> @llvm.spv.faceforward.v3f32(<3 x float> %{{.*}}, <3 x float> %{{.*}}, <3 x float> %{{.*}})
-// SPVCHECK: ret <3 x float> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fdot.v3f32(<3 x float> %{{.*}}, <3 x float> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <3 x float> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x float> %{{.*}}, <3 x float> %fneg.i
+// SPVCHECK: ret <3 x float> %hlsl.select.i
float3 test_faceforward_float3(float3 N, float3 I, float3 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float4
@@ -89,6 +110,9 @@ float3 test_faceforward_float3(float3 N, float3 I, float3 Ng) { return faceforwa
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x float> %{{.*}}, <4 x float> %fneg.i
// CHECK: ret <4 x float> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float4
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <4 x float> @llvm.spv.faceforward.v4f32(<4 x float> %{{.*}}, <4 x float> %{{.*}}, <4 x float> %{{.*}})
-// SPVCHECK: ret <4 x float> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fdot.v4f32(<4 x float> %{{.*}}, <4 x float> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <4 x float> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x float> %{{.*}}, <4 x float> %fneg.i
+// SPVCHECK: ret <4 x float> %hlsl.select.i
float4 test_faceforward_float4(float4 N, float4 I, float4 Ng) { return faceforward(N, I, Ng); }
diff --git a/llvm/lib/Target/SPIRV/SPIRVCombine.td b/llvm/lib/Target/SPIRV/SPIRVCombine.td
index 6f726e024de52..78a57146e61d2 100644
--- a/llvm/lib/Target/SPIRV/SPIRVCombine.td
+++ b/llvm/lib/Target/SPIRV/SPIRVCombine.td
@@ -15,8 +15,15 @@ def vector_length_sub_to_distance_lowering : GICombineRule <
(apply [{ applySPIRVDistance(*${root}, MRI, B); }])
>;
+def vector_select_to_faceforward_lowering : GICombineRule <
+ (defs root:$root),
+ (match (wip_match_opcode G_SELECT):$root,
+ [{ return matchSelectToFaceForward(*${root}, MRI); }]),
+ (apply [{ applySPIRVFaceForward(*${root}, MRI, B); }])
+>;
+
def SPIRVPreLegalizerCombiner
: GICombiner<"SPIRVPreLegalizerCombinerImpl",
- [vector_length_sub_to_distance_lowering]> {
+ [vector_length_sub_to_distance_lowering, vector_select_to_faceforward_lowering]> {
let CombineAllMethodName = "tryCombineAllImpl";
}
diff --git a/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp b/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp
index c96ee6b02491a..82c5cbf75b849 100644
--- a/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp
@@ -50,6 +50,18 @@ namespace {
#include "SPIRVGenPreLegalizeGICombiner.inc"
#undef GET_GICOMBINER_TYPES
+static void removeAllUses(Register Reg, MachineRegisterInfo &MRI,
+ SPIRVGlobalRegistry *GR) {
+ SmallVector<MachineInstr *, 4> UsesToErase(
+ llvm::make_pointer_range(MRI.use_instructions(Reg)));
+
+ // calling eraseFromParent too early invalidates the iterator.
+ for (auto *MIToErase : UsesToErase) {
+ GR->invalidateMachineInstr(MIToErase);
+ MIToErase->eraseFromParent();
+ }
+}
+
/// This match is part of a combine that
/// rewrites length(X - Y) to distance(X, Y)
/// (f32 (g_intrinsic length
@@ -98,21 +110,98 @@ void applySPIRVDistance(MachineInstr &MI, MachineRegisterInfo &MRI,
SPIRVGlobalRegistry *GR =
MI.getMF()->getSubtarget<SPIRVSubtarget>().getSPIRVGlobalRegistry();
- auto RemoveAllUses = [&](Register Reg) {
- SmallVector<MachineInstr *, 4> UsesToErase(
- llvm::make_pointer_range(MRI.use_instructions(Reg)));
-
- // calling eraseFromParent to early invalidates the iterator.
- for (auto *MIToErase : UsesToErase) {
- GR->invalidateMachineInstr(MIToErase);
- MIToErase->eraseFromParent();
- }
- };
- RemoveAllUses(SubDestReg); // remove all uses of FSUB Result
+ removeAllUses(SubDestReg, MRI, GR); // remove all uses of FSUB Result
GR->invalidateMachineInstr(SubInstr);
SubInstr->eraseFromParent(); // remove FSUB instruction
}
+/// This match is part of a combine that
+/// rewrites select(fcmp(dot(I, Ng), 0), N, 0 - N) to faceforward(N, I, Ng)
+/// (vXf32 (g_select
+/// (g_fcmp
+/// (g_intrinsic dot(vXf32 I) (vXf32 Ng)
+/// 0)
+/// (vXf32 N)
+/// (vXf32 g_fsub (0) (vXf32 N))))
+/// ->
+/// (vXf32 (g_intrinsic faceforward
+/// (vXf32 N) (vXf32 I) (vXf32 Ng)))
+///
+bool matchSelectToFaceForward(MachineInstr &MI, MachineRegisterInfo &MRI) {
+ if (MI.getOpcode() != TargetOpcode::G_SELECT)
+ return false;
+
+ // Check if select's condition is a comparison between a dot product and 0.
+ Register CondReg = MI.getOperand(1).getReg();
+ MachineInstr *CondInstr = MRI.getVRegDef(CondReg);
+ if (!CondInstr || CondInstr->getOpcode() != TargetOpcode::G_FCMP)
+ return false;
+
+ Register DotReg = CondInstr->getOperand(2).getReg();
+ MachineInstr *DotInstr = MRI.getVRegDef(DotReg);
+ if (DotInstr->getOpcode() != TargetOpcode::G_FMUL &&
+ (DotInstr->getOpcode() != TargetOpcode::G_INTRINSIC ||
+ cast<GIntrinsic>(DotInstr)->getIntrinsicID() != Intrinsic::spv_fdot))
+ return false;
+
+ Register CondZeroReg = CondInstr->getOperand(3).getReg();
+ MachineInstr *CondZeroInstr = MRI.getVRegDef(CondZeroReg);
+ if (CondZeroInstr->getOpcode() != TargetOpcode::G_FCONSTANT ||
+ !CondZeroInstr->getOperand(1).getFPImm()->isZero())
+ return false;
+
+ // Check if select's false operand is the negation of the true operand.
+ Register TrueReg = MI.getOperand(2).getReg();
+ Register FalseReg = MI.getOperand(3).getReg();
+ MachineInstr *FalseInstr = MRI.getVRegDef(FalseReg);
+ if (FalseInstr->getOpcode() != TargetOpcode::G_FNEG)
+ return false;
+ if (TrueReg != FalseInstr->getOperand(1).getReg())
+ return false;
+
+ return true;
+}
+void applySPIRVFaceForward(MachineInstr &MI, MachineRegisterInfo &MRI,
+ MachineIRBuilder &B) {
+
+ // Extract the operands for N, I, and Ng from the match criteria.
+ Register CondReg = MI.getOperand(1).getReg();
+ MachineInstr *CondInstr = MRI.getVRegDef(CondReg);
+ Register DotReg = CondInstr->getOperand(2).getReg();
+ MachineInstr *DotInstr = MRI.getVRegDef(DotReg);
+ Register DotOperand1, DotOperand2;
+ if (DotInstr->getOpcode() == TargetOpcode::G_FMUL) {
+ DotOperand1 = DotInstr->getOperand(1).getReg();
+ DotOperand2 = DotInstr->getOperand(2).getReg();
+ } else {
+ DotOperand1 = DotInstr->getOperand(2).getReg();
+ DotOperand2 = DotInstr->getOperand(3).getReg();
+ }
+ Register TrueReg = MI.getOperand(2).getReg();
+
+ // Remove the original `select` instruction.
+ Register ResultReg = MI.getOperand(0).getReg();
+ DebugLoc DL = MI.getDebugLoc();
+ MachineBasicBlock &MBB = *MI.getParent();
+ MachineBasicBlock::iterator InsertPt = MI.getIterator();
+
+ // Build the `spv_faceforward` intrinsic.
+ MachineInstrBuilder NewInstr =
+ BuildMI(MBB, InsertPt, DL, B.getTII().get(TargetOpcode::G_INTRINSIC));
+ NewInstr
+ .addDef(ResultReg) // Result register
+ .addIntrinsicID(Intrinsic::spv_faceforward) // Intrinsic ID
+ .addUse(TrueReg) // Operand N
+ .addUse(DotOperand1) // Operand I
+ .addUse(DotOperand2); // Operand Ng
+
+ SPIRVGlobalRegistry *GR =
+ MI.getMF()->getSubtarget<SPIRVSubtarget>().getSPIRVGlobalRegistry();
+ removeAllUses(CondReg, MRI, GR); // remove all uses of FCMP Result
+ GR->invalidateMachineInstr(CondInstr);
+ CondInstr->eraseFromParent(); // remove FCMP instruction
+}
+
class SPIRVPreLegalizerCombinerImpl : public Combiner {
protected:
const CombinerHelper Helper;
diff --git a/llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir b/llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
new file mode 100644
index 0000000000000..08f03942460ba
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
@@ -0,0 +1,63 @@
+# RUN: llc -verify-machineinstrs -O0 -mtriple spirv-unknown-unknown -run-pass=spirv-prelegalizer-combiner %s -o - | FileCheck %s
+# REQUIRES: asserts
+---
+name: faceforward_instcombine_float
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.1.entry:
+ ; CHECK-LABEL: name: faceforward_instcombine_float
+ ; CHECK-NOT: %9:_(s32) = G_FCONSTANT float 0.000000e+00
+ ; CHECK-NOT: %8:_(s32) = G_FMUL %1:fid, %2:fid
+ ; CHECK-NOT: %10:_(s1) = G_FCMP floatpred(olt), %8:_(s32), %9:_
+ ; CHECK-NOT: %11:_(s32) = G_FNEG %0:fid
+ ; CHECK-NOT: %12:id(s32) = G_SELECT %10:_(s1), %0:fid, %11:_
+ ; CHECK: %10:id(s32) = G_INTRINSIC intrinsic(@llvm.spv.faceforward), %2(s32), %3(s32), %4(s32)
+ %3:type(s64) = OpTypeFloat 32
+ %5:type(s64) = OpTypeFunction %3:type(s64), %3:type(s64), %3:type(s64), %3:type(s64)
+ OpName %0:fid(s32), 97
+ OpName %1:fid(s32), 98
+ OpName %2:fid(s32), 99
+ %4:iid(s64) = OpFunction %3:type(s64), 0, %5:type(s64)
+ %0:fid(s32) = OpFunctionParameter %3:type(s64)
+ %1:fid(s32) = OpFunctionParameter %3:type(s64)
+ %2:fid(s32) = OpFunctionParameter %3:type(s64)
+ OpName %4:iid(s64), 1701011814, 2003988326, 1600418401, 1953721961, 1651339107, 1600482921, 1634692198, 116
+ %9:_(s32) = G_FCONSTANT float 0.000000e+00
+ %8:_(s32) = G_FMUL %1:fid, %2:fid
+ %10:_(s1) = G_FCMP floatpred(olt), %8:_(s32), %9:_
+ %11:_(s32) = G_FNEG %0:fid
+ %12:id(s32) = G_SELECT %10:_(s1), %0:fid, %11:_
+ OpReturnValue %12:id(s32)
+---
+name: faceforward_instcombine_float4
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.1.entry:
+ ; CHECK-LABEL: name: faceforward_instcombine_float4
+ ; CHECK-NOT: %10:_(s32) = G_FCONSTANT float 0.000000e+00
+ ; CHECK-NOT: %9:_(s32) = G_INTRINSIC intrinsic(@llvm.spv.fdot), %1:vfid(<4 x s32>), %2:vfid(<4 x s32>)
+ ; CHECK-NOT: %11:_(s1) = G_FCMP floatpred(olt), %9:_(s32), %10:_
+ ; CHECK-NOT: %12:_(<4 x s32>) = G_FNEG %0:vfid
+ ; CHECK-NOT: %13:id(<4 x s32>) = G_SELECT %11:_(s1), %0:vfid, %12:_
+ ; CHECK: %11:id(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.spv.faceforward), %3(<4 x s32>), %4(<4 x s32>), %5(<4 x s32>)
+ %4:type(s64) = OpTypeVector %3:type(s64), 4
+ %6:type(s64) = OpTypeFunction %4:type(s64), %4:type(s64), %4:type(s64), %4:type(s64)
+ %3:type(s64) = OpTypeFloat 32
+ OpName %0:vfid(<4 x s32>), 97
+ OpName %1:vfid(<4 x s32>), 98
+ OpName %2:vfid(<4 x s32>), 99
+ %5:iid(s64) = OpFunction %4:type(s64), 0, %6:type(s64)
+ %0:vfid(<4 x s32>) = OpFunctionParameter %4:type(s64)
+ %1:vfid(<4 x s32>) = OpFunctionParameter %4:type(s64)
+ %2:vfid(<4 x s32>) = OpFunctionParameter %4:type(s64)
+ OpName %5:iid(s64), 1701011814, 2003988326, 1600418401, 1953721961, 1651339107, 1600482921, 1634692198, 13428
+ OpDecorate %5:iid(s64), 41, 1701011814, 2003988326, 1600418401, 1953721961, 1651339107, 1600482921, 1634692198, 13428, 0
+ %10:_(s32) = G_FCONSTANT float 0.000000e+00
+ %9:_(s32) = G_INTRINSIC intrinsic(@llvm.spv.fdot), %1:vfid(<4 x s32>), %2:vfid(<4 x s32>)
+ %11:_(s1) = G_FCMP floatpred(olt), %9:_(s32), %10:_
+ %12:_(<4 x s32>) = G_FNEG %0:vfid
+ %13:id(<4 x s32>) = G_SELECT %11:_(s1), %0:vfid, %12:_
+ OpReturnValue %13:id(<4 x s32>)
+
\ No newline at end of file
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll
index 22742169506f2..6811f95bbab17 100644
--- a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll
@@ -1,7 +1,5 @@
-; RUN: llc -O0 -verify-machineinstrs -mtriple=spirv-unknown-unknown %s -o - | FileCheck %s
-; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-unknown-unknown %s -o - -filetype=obj | spirv-val --target-env spv1.4 %}
-
-; FIXME(#136344): Change --target-env to vulkan1.3 and update this test accordingly once the issue is resolved.
+; RUN: llc -O0 -verify-machineinstrs -mtriple=spirv-vulkan1.3-unknown %s -o...
[truncated]
|
Member
|
@llvm/pr-subscribers-hlsl Author: Kaitlin Peng (kmpeng) ChangesTasks completed:
Closes #137255. Patch is 25.07 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/139959.diff 7 Files Affected:
diff --git a/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h b/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h
index 4eb7b8f45c85a..e1996fd3e30ed 100644
--- a/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h
+++ b/clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h
@@ -127,11 +127,7 @@ template <typename T> constexpr vector<T, 4> lit_impl(T NDotL, T NDotH, T M) {
}
template <typename T> constexpr T faceforward_impl(T N, T I, T Ng) {
-#if (__has_builtin(__builtin_spirv_faceforward))
- return __builtin_spirv_faceforward(N, I, Ng);
-#else
return select(dot(I, Ng) < 0, N, -N);
-#endif
}
template <typename T> constexpr T ldexp_impl(T X, T Exp) {
diff --git a/clang/test/CodeGenHLSL/builtins/faceforward.hlsl b/clang/test/CodeGenHLSL/builtins/faceforward.hlsl
index d2ece57aba4ae..f12f75f384964 100644
--- a/clang/test/CodeGenHLSL/builtins/faceforward.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/faceforward.hlsl
@@ -12,8 +12,11 @@
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, half %{{.*}}, half %fneg.i
// CHECK: ret half %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef half @llvm.spv.faceforward.f16(half %{{.*}}, half %{{.*}}, half %{{.*}})
-// SPVCHECK: ret half %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = fmul reassoc nnan ninf nsz arcp afn half %{{.*}}, %{{.*}}
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn half %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, half %{{.*}}, half %fneg.i
+// SPVCHECK: ret half %hlsl.select.i
half test_faceforward_half(half N, half I, half Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_half2
@@ -23,8 +26,11 @@ half test_faceforward_half(half N, half I, half Ng) { return faceforward(N, I, N
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x half> %{{.*}}, <2 x half> %fneg.i
// CHECK: ret <2 x half> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half2
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <2 x half> @llvm.spv.faceforward.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, <2 x half> %{{.*}})
-// SPVCHECK: ret <2 x half> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <2 x half> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x half> %{{.*}}, <2 x half> %fneg.i
+// SPVCHECK: ret <2 x half> %hlsl.select.i
half2 test_faceforward_half2(half2 N, half2 I, half2 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_half3
@@ -34,8 +40,11 @@ half2 test_faceforward_half2(half2 N, half2 I, half2 Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x half> %{{.*}}, <3 x half> %fneg.i
// CHECK: ret <3 x half> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half3
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <3 x half> @llvm.spv.faceforward.v3f16(<3 x half> %{{.*}}, <3 x half> %{{.*}}, <3 x half> %{{.*}})
-// SPVCHECK: ret <3 x half> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v3f16(<3 x half> %{{.*}}, <3 x half> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <3 x half> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x half> %{{.*}}, <3 x half> %fneg.i
+// SPVCHECK: ret <3 x half> %hlsl.select.i
half3 test_faceforward_half3(half3 N, half3 I, half3 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_half4
@@ -45,8 +54,11 @@ half3 test_faceforward_half3(half3 N, half3 I, half3 Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x half> %{{.*}}, <4 x half> %fneg.i
// CHECK: ret <4 x half> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_half4
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <4 x half> @llvm.spv.faceforward.v4f16(<4 x half> %{{.*}}, <4 x half> %{{.*}}, <4 x half> %{{.*}})
-// SPVCHECK: ret <4 x half> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v4f16(<4 x half> %{{.*}}, <4 x half> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt half %hlsl.dot.i, 0xH0000
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <4 x half> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x half> %{{.*}}, <4 x half> %fneg.i
+// SPVCHECK: ret <4 x half> %hlsl.select.i
half4 test_faceforward_half4(half4 N, half4 I, half4 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float
@@ -56,8 +68,11 @@ half4 test_faceforward_half4(half4 N, half4 I, half4 Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, float %{{.*}}, float %fneg.i
// CHECK: ret float %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef float @llvm.spv.faceforward.f32(float %{{.*}}, float %{{.*}}, float %{{.*}})
-// SPVCHECK: ret float %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = fmul reassoc nnan ninf nsz arcp afn float %{{.*}}, %{{.*}}
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn float %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, float %{{.*}}, float %fneg.i
+// SPVCHECK: ret float %hlsl.select.i
float test_faceforward_float(float N, float I, float Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float2
@@ -67,8 +82,11 @@ float test_faceforward_float(float N, float I, float Ng) { return faceforward(N,
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x float> %{{.*}}, <2 x float> %fneg.i
// CHECK: ret <2 x float> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float2
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <2 x float> @llvm.spv.faceforward.v2f32(<2 x float> %{{.*}}, <2 x float> %{{.*}}, <2 x float> %{{.*}})
-// SPVCHECK: ret <2 x float> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fdot.v2f32(<2 x float> %{{.*}}, <2 x float> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <2 x float> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <2 x float> %{{.*}}, <2 x float> %fneg.i
+// SPVCHECK: ret <2 x float> %hlsl.select.i
float2 test_faceforward_float2(float2 N, float2 I, float2 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float3
@@ -78,8 +96,11 @@ float2 test_faceforward_float2(float2 N, float2 I, float2 Ng) { return faceforwa
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x float> %{{.*}}, <3 x float> %fneg.i
// CHECK: ret <3 x float> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float3
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <3 x float> @llvm.spv.faceforward.v3f32(<3 x float> %{{.*}}, <3 x float> %{{.*}}, <3 x float> %{{.*}})
-// SPVCHECK: ret <3 x float> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fdot.v3f32(<3 x float> %{{.*}}, <3 x float> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <3 x float> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <3 x float> %{{.*}}, <3 x float> %fneg.i
+// SPVCHECK: ret <3 x float> %hlsl.select.i
float3 test_faceforward_float3(float3 N, float3 I, float3 Ng) { return faceforward(N, I, Ng); }
// CHECK-LABEL: test_faceforward_float4
@@ -89,6 +110,9 @@ float3 test_faceforward_float3(float3 N, float3 I, float3 Ng) { return faceforwa
// CHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x float> %{{.*}}, <4 x float> %fneg.i
// CHECK: ret <4 x float> %hlsl.select.i
// SPVCHECK-LABEL: test_faceforward_float4
-// SPVCHECK: %spv.faceforward.i = call reassoc nnan ninf nsz arcp afn noundef <4 x float> @llvm.spv.faceforward.v4f32(<4 x float> %{{.*}}, <4 x float> %{{.*}}, <4 x float> %{{.*}})
-// SPVCHECK: ret <4 x float> %spv.faceforward.i
+// SPVCHECK: %hlsl.dot.i = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fdot.v4f32(<4 x float> %{{.*}}, <4 x float> %{{.*}})
+// SPVCHECK: %cmp.i = fcmp reassoc nnan ninf nsz arcp afn olt float %hlsl.dot.i, 0.000000e+00
+// SPVCHECK: %fneg.i = fneg reassoc nnan ninf nsz arcp afn <4 x float> %{{.*}}
+// SPVCHECK: %hlsl.select.i = select reassoc nnan ninf nsz arcp afn i1 %cmp.i, <4 x float> %{{.*}}, <4 x float> %fneg.i
+// SPVCHECK: ret <4 x float> %hlsl.select.i
float4 test_faceforward_float4(float4 N, float4 I, float4 Ng) { return faceforward(N, I, Ng); }
diff --git a/llvm/lib/Target/SPIRV/SPIRVCombine.td b/llvm/lib/Target/SPIRV/SPIRVCombine.td
index 6f726e024de52..78a57146e61d2 100644
--- a/llvm/lib/Target/SPIRV/SPIRVCombine.td
+++ b/llvm/lib/Target/SPIRV/SPIRVCombine.td
@@ -15,8 +15,15 @@ def vector_length_sub_to_distance_lowering : GICombineRule <
(apply [{ applySPIRVDistance(*${root}, MRI, B); }])
>;
+def vector_select_to_faceforward_lowering : GICombineRule <
+ (defs root:$root),
+ (match (wip_match_opcode G_SELECT):$root,
+ [{ return matchSelectToFaceForward(*${root}, MRI); }]),
+ (apply [{ applySPIRVFaceForward(*${root}, MRI, B); }])
+>;
+
def SPIRVPreLegalizerCombiner
: GICombiner<"SPIRVPreLegalizerCombinerImpl",
- [vector_length_sub_to_distance_lowering]> {
+ [vector_length_sub_to_distance_lowering, vector_select_to_faceforward_lowering]> {
let CombineAllMethodName = "tryCombineAllImpl";
}
diff --git a/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp b/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp
index c96ee6b02491a..82c5cbf75b849 100644
--- a/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVPreLegalizerCombiner.cpp
@@ -50,6 +50,18 @@ namespace {
#include "SPIRVGenPreLegalizeGICombiner.inc"
#undef GET_GICOMBINER_TYPES
+static void removeAllUses(Register Reg, MachineRegisterInfo &MRI,
+ SPIRVGlobalRegistry *GR) {
+ SmallVector<MachineInstr *, 4> UsesToErase(
+ llvm::make_pointer_range(MRI.use_instructions(Reg)));
+
+ // calling eraseFromParent too early invalidates the iterator.
+ for (auto *MIToErase : UsesToErase) {
+ GR->invalidateMachineInstr(MIToErase);
+ MIToErase->eraseFromParent();
+ }
+}
+
/// This match is part of a combine that
/// rewrites length(X - Y) to distance(X, Y)
/// (f32 (g_intrinsic length
@@ -98,21 +110,98 @@ void applySPIRVDistance(MachineInstr &MI, MachineRegisterInfo &MRI,
SPIRVGlobalRegistry *GR =
MI.getMF()->getSubtarget<SPIRVSubtarget>().getSPIRVGlobalRegistry();
- auto RemoveAllUses = [&](Register Reg) {
- SmallVector<MachineInstr *, 4> UsesToErase(
- llvm::make_pointer_range(MRI.use_instructions(Reg)));
-
- // calling eraseFromParent to early invalidates the iterator.
- for (auto *MIToErase : UsesToErase) {
- GR->invalidateMachineInstr(MIToErase);
- MIToErase->eraseFromParent();
- }
- };
- RemoveAllUses(SubDestReg); // remove all uses of FSUB Result
+ removeAllUses(SubDestReg, MRI, GR); // remove all uses of FSUB Result
GR->invalidateMachineInstr(SubInstr);
SubInstr->eraseFromParent(); // remove FSUB instruction
}
+/// This match is part of a combine that
+/// rewrites select(fcmp(dot(I, Ng), 0), N, 0 - N) to faceforward(N, I, Ng)
+/// (vXf32 (g_select
+/// (g_fcmp
+/// (g_intrinsic dot(vXf32 I) (vXf32 Ng)
+/// 0)
+/// (vXf32 N)
+/// (vXf32 g_fsub (0) (vXf32 N))))
+/// ->
+/// (vXf32 (g_intrinsic faceforward
+/// (vXf32 N) (vXf32 I) (vXf32 Ng)))
+///
+bool matchSelectToFaceForward(MachineInstr &MI, MachineRegisterInfo &MRI) {
+ if (MI.getOpcode() != TargetOpcode::G_SELECT)
+ return false;
+
+ // Check if select's condition is a comparison between a dot product and 0.
+ Register CondReg = MI.getOperand(1).getReg();
+ MachineInstr *CondInstr = MRI.getVRegDef(CondReg);
+ if (!CondInstr || CondInstr->getOpcode() != TargetOpcode::G_FCMP)
+ return false;
+
+ Register DotReg = CondInstr->getOperand(2).getReg();
+ MachineInstr *DotInstr = MRI.getVRegDef(DotReg);
+ if (DotInstr->getOpcode() != TargetOpcode::G_FMUL &&
+ (DotInstr->getOpcode() != TargetOpcode::G_INTRINSIC ||
+ cast<GIntrinsic>(DotInstr)->getIntrinsicID() != Intrinsic::spv_fdot))
+ return false;
+
+ Register CondZeroReg = CondInstr->getOperand(3).getReg();
+ MachineInstr *CondZeroInstr = MRI.getVRegDef(CondZeroReg);
+ if (CondZeroInstr->getOpcode() != TargetOpcode::G_FCONSTANT ||
+ !CondZeroInstr->getOperand(1).getFPImm()->isZero())
+ return false;
+
+ // Check if select's false operand is the negation of the true operand.
+ Register TrueReg = MI.getOperand(2).getReg();
+ Register FalseReg = MI.getOperand(3).getReg();
+ MachineInstr *FalseInstr = MRI.getVRegDef(FalseReg);
+ if (FalseInstr->getOpcode() != TargetOpcode::G_FNEG)
+ return false;
+ if (TrueReg != FalseInstr->getOperand(1).getReg())
+ return false;
+
+ return true;
+}
+void applySPIRVFaceForward(MachineInstr &MI, MachineRegisterInfo &MRI,
+ MachineIRBuilder &B) {
+
+ // Extract the operands for N, I, and Ng from the match criteria.
+ Register CondReg = MI.getOperand(1).getReg();
+ MachineInstr *CondInstr = MRI.getVRegDef(CondReg);
+ Register DotReg = CondInstr->getOperand(2).getReg();
+ MachineInstr *DotInstr = MRI.getVRegDef(DotReg);
+ Register DotOperand1, DotOperand2;
+ if (DotInstr->getOpcode() == TargetOpcode::G_FMUL) {
+ DotOperand1 = DotInstr->getOperand(1).getReg();
+ DotOperand2 = DotInstr->getOperand(2).getReg();
+ } else {
+ DotOperand1 = DotInstr->getOperand(2).getReg();
+ DotOperand2 = DotInstr->getOperand(3).getReg();
+ }
+ Register TrueReg = MI.getOperand(2).getReg();
+
+ // Remove the original `select` instruction.
+ Register ResultReg = MI.getOperand(0).getReg();
+ DebugLoc DL = MI.getDebugLoc();
+ MachineBasicBlock &MBB = *MI.getParent();
+ MachineBasicBlock::iterator InsertPt = MI.getIterator();
+
+ // Build the `spv_faceforward` intrinsic.
+ MachineInstrBuilder NewInstr =
+ BuildMI(MBB, InsertPt, DL, B.getTII().get(TargetOpcode::G_INTRINSIC));
+ NewInstr
+ .addDef(ResultReg) // Result register
+ .addIntrinsicID(Intrinsic::spv_faceforward) // Intrinsic ID
+ .addUse(TrueReg) // Operand N
+ .addUse(DotOperand1) // Operand I
+ .addUse(DotOperand2); // Operand Ng
+
+ SPIRVGlobalRegistry *GR =
+ MI.getMF()->getSubtarget<SPIRVSubtarget>().getSPIRVGlobalRegistry();
+ removeAllUses(CondReg, MRI, GR); // remove all uses of FCMP Result
+ GR->invalidateMachineInstr(CondInstr);
+ CondInstr->eraseFromParent(); // remove FCMP instruction
+}
+
class SPIRVPreLegalizerCombinerImpl : public Combiner {
protected:
const CombinerHelper Helper;
diff --git a/llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir b/llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
new file mode 100644
index 0000000000000..08f03942460ba
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
@@ -0,0 +1,63 @@
+# RUN: llc -verify-machineinstrs -O0 -mtriple spirv-unknown-unknown -run-pass=spirv-prelegalizer-combiner %s -o - | FileCheck %s
+# REQUIRES: asserts
+---
+name: faceforward_instcombine_float
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.1.entry:
+ ; CHECK-LABEL: name: faceforward_instcombine_float
+ ; CHECK-NOT: %9:_(s32) = G_FCONSTANT float 0.000000e+00
+ ; CHECK-NOT: %8:_(s32) = G_FMUL %1:fid, %2:fid
+ ; CHECK-NOT: %10:_(s1) = G_FCMP floatpred(olt), %8:_(s32), %9:_
+ ; CHECK-NOT: %11:_(s32) = G_FNEG %0:fid
+ ; CHECK-NOT: %12:id(s32) = G_SELECT %10:_(s1), %0:fid, %11:_
+ ; CHECK: %10:id(s32) = G_INTRINSIC intrinsic(@llvm.spv.faceforward), %2(s32), %3(s32), %4(s32)
+ %3:type(s64) = OpTypeFloat 32
+ %5:type(s64) = OpTypeFunction %3:type(s64), %3:type(s64), %3:type(s64), %3:type(s64)
+ OpName %0:fid(s32), 97
+ OpName %1:fid(s32), 98
+ OpName %2:fid(s32), 99
+ %4:iid(s64) = OpFunction %3:type(s64), 0, %5:type(s64)
+ %0:fid(s32) = OpFunctionParameter %3:type(s64)
+ %1:fid(s32) = OpFunctionParameter %3:type(s64)
+ %2:fid(s32) = OpFunctionParameter %3:type(s64)
+ OpName %4:iid(s64), 1701011814, 2003988326, 1600418401, 1953721961, 1651339107, 1600482921, 1634692198, 116
+ %9:_(s32) = G_FCONSTANT float 0.000000e+00
+ %8:_(s32) = G_FMUL %1:fid, %2:fid
+ %10:_(s1) = G_FCMP floatpred(olt), %8:_(s32), %9:_
+ %11:_(s32) = G_FNEG %0:fid
+ %12:id(s32) = G_SELECT %10:_(s1), %0:fid, %11:_
+ OpReturnValue %12:id(s32)
+---
+name: faceforward_instcombine_float4
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.1.entry:
+ ; CHECK-LABEL: name: faceforward_instcombine_float4
+ ; CHECK-NOT: %10:_(s32) = G_FCONSTANT float 0.000000e+00
+ ; CHECK-NOT: %9:_(s32) = G_INTRINSIC intrinsic(@llvm.spv.fdot), %1:vfid(<4 x s32>), %2:vfid(<4 x s32>)
+ ; CHECK-NOT: %11:_(s1) = G_FCMP floatpred(olt), %9:_(s32), %10:_
+ ; CHECK-NOT: %12:_(<4 x s32>) = G_FNEG %0:vfid
+ ; CHECK-NOT: %13:id(<4 x s32>) = G_SELECT %11:_(s1), %0:vfid, %12:_
+ ; CHECK: %11:id(<4 x s32>) = G_INTRINSIC intrinsic(@llvm.spv.faceforward), %3(<4 x s32>), %4(<4 x s32>), %5(<4 x s32>)
+ %4:type(s64) = OpTypeVector %3:type(s64), 4
+ %6:type(s64) = OpTypeFunction %4:type(s64), %4:type(s64), %4:type(s64), %4:type(s64)
+ %3:type(s64) = OpTypeFloat 32
+ OpName %0:vfid(<4 x s32>), 97
+ OpName %1:vfid(<4 x s32>), 98
+ OpName %2:vfid(<4 x s32>), 99
+ %5:iid(s64) = OpFunction %4:type(s64), 0, %6:type(s64)
+ %0:vfid(<4 x s32>) = OpFunctionParameter %4:type(s64)
+ %1:vfid(<4 x s32>) = OpFunctionParameter %4:type(s64)
+ %2:vfid(<4 x s32>) = OpFunctionParameter %4:type(s64)
+ OpName %5:iid(s64), 1701011814, 2003988326, 1600418401, 1953721961, 1651339107, 1600482921, 1634692198, 13428
+ OpDecorate %5:iid(s64), 41, 1701011814, 2003988326, 1600418401, 1953721961, 1651339107, 1600482921, 1634692198, 13428, 0
+ %10:_(s32) = G_FCONSTANT float 0.000000e+00
+ %9:_(s32) = G_INTRINSIC intrinsic(@llvm.spv.fdot), %1:vfid(<4 x s32>), %2:vfid(<4 x s32>)
+ %11:_(s1) = G_FCMP floatpred(olt), %9:_(s32), %10:_
+ %12:_(<4 x s32>) = G_FNEG %0:vfid
+ %13:id(<4 x s32>) = G_SELECT %11:_(s1), %0:vfid, %12:_
+ OpReturnValue %13:id(<4 x s32>)
+
\ No newline at end of file
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll
index 22742169506f2..6811f95bbab17 100644
--- a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/faceforward.ll
@@ -1,7 +1,5 @@
-; RUN: llc -O0 -verify-machineinstrs -mtriple=spirv-unknown-unknown %s -o - | FileCheck %s
-; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-unknown-unknown %s -o - -filetype=obj | spirv-val --target-env spv1.4 %}
-
-; FIXME(#136344): Change --target-env to vulkan1.3 and update this test accordingly once the issue is resolved.
+; RUN: llc -O0 -verify-machineinstrs -mtriple=spirv-vulkan1.3-unknown %s -o...
[truncated]
|
arsenm
reviewed
May 14, 2025
Contributor
arsenm
left a comment
arsenm
left a comment
There was a problem hiding this comment.
Description mentions instcombine but this has nothing to do with instcombine
farzonl
reviewed
May 15, 2025
farzonl
reviewed
May 15, 2025
llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
Outdated
Show resolved
Hide resolved
farzonl
reviewed
May 15, 2025
llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
Show resolved
Hide resolved
farzonl
reviewed
May 15, 2025
llvm/test/CodeGen/SPIRV/GlobalISel/InstCombine/prelegalizercombiner-select-to-faceforward.mir
Show resolved
Hide resolved
farzonl
reviewed
May 15, 2025
arsenm
reviewed
May 16, 2025
Contributor
arsenm
left a comment
arsenm
left a comment
There was a problem hiding this comment.
Description should not mention instcombine, instcombine is not related
s-perron
reviewed
May 16, 2025
Contributor
s-perron
left a comment
s-perron
left a comment
There was a problem hiding this comment.
I have a minor concern about how specific the N and 0-N checks are. Otherwise this looks good to me.
arsenm
reviewed
May 16, 2025
kmpeng
changed the title
faceforwardfaceforward
kmpeng
force-pushed
the
faceforward-patternmatch
branch
from
5cb68c5 to
6d19a11
Compare
kmpeng
force-pushed
the
faceforward-patternmatch
branch
2 times, most recently
from
67cad12 to
6868f35
Compare
arsenm
reviewed
Sep 25, 2025
kmpeng
force-pushed
the
faceforward-patternmatch
branch
from
64df069 to
900b7bb
Compare
farzonl
reviewed
Oct 17, 2025
farzonl
reviewed
Oct 17, 2025
farzonl
reviewed
Oct 17, 2025
farzonl
reviewed
Oct 17, 2025
farzonl
reviewed
Oct 17, 2025
kmpeng
force-pushed
the
faceforward-patternmatch
branch
from
4871e02 to
824225d
Compare
farzonl
approved these changes
Nov 4, 2025
Member
farzonl
left a comment
farzonl
left a comment
There was a problem hiding this comment.
LGTM, wait for other feedback though.
kmpeng
force-pushed
the
faceforward-patternmatch
branch
from
824225d to
eb3994f
Compare
Contributor
Author
🐧 Linux x64 Test Results
|
augusto2112
pushed a commit
to augusto2112/llvm-project
that referenced
this pull request
Dec 3, 2025
) Tasks completed: - Pattern match`select(fcmp(dot(p2, p3), 0), p1, -p1)` to `faceforward(p1, p2, p3)` - Add pattern matching tests to `prelegalizercombiner-select-to-faceforward.mir` and `faceforward.ll` - Add CL extension error test `llvm/test/CodeGen/SPIRV/opencl/faceforward-error.ll` - Add CL extension test for no pattern matching in `llvm/test/CodeGen/SPIRV/opencl/faceforward.ll` Closes llvm#137255.
kcloudy0717
pushed a commit
to kcloudy0717/llvm-project
that referenced
this pull request
Dec 4, 2025
) Tasks completed: - Pattern match`select(fcmp(dot(p2, p3), 0), p1, -p1)` to `faceforward(p1, p2, p3)` - Add pattern matching tests to `prelegalizercombiner-select-to-faceforward.mir` and `faceforward.ll` - Add CL extension error test `llvm/test/CodeGen/SPIRV/opencl/faceforward-error.ll` - Add CL extension test for no pattern matching in `llvm/test/CodeGen/SPIRV/opencl/faceforward.ll` Closes llvm#137255.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Tasks completed:
select(fcmp(dot(p2, p3), 0), p1, -p1)tofaceforward(p1, p2, p3)prelegalizercombiner-select-to-faceforward.mirandfaceforward.llllvm/test/CodeGen/SPIRV/opencl/faceforward-error.llllvm/test/CodeGen/SPIRV/opencl/faceforward.llCloses #137255.