[InstCombine] CVTPH2PS Vector Demanded Elements + Constant Folding

[RKSimon]

Summary

Improved InstCombine support for CVTPH2PS (F16C half 2 float conversion):

<4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16>) - only uses the bottom 4 i16 elements for the conversion.

Added constant folding support.

[ab]

LGTM

lib/Transforms/InstCombine/InstCombineCalls.cpp

846–857 ↗

(On Diff #34338)

I couldn't help but notice that this is very similar to the ppc case above. All else being equal, can you keep them together?

Plus, the ph2ps case is similar to ss2si below!

test/Transforms/InstCombine/x86-f16c.ll

3–4 ↗

(On Diff #34338)

Unnecessary attributes?

[RKSimon]

Thanks Ahmed, I'll look into doing something similar for CVTPS2PH in a future patch, although I'm a little concerned about matching all the (compile-time) rounding modes.

lib/Transforms/InstCombine/InstCombineCalls.cpp

846–857 ↗

(On Diff #34338)

Yes there is a lot of code duplication going in this file - splitting off common code into helper functions should be quite straightforward.

test/Transforms/InstCombine/x86-f16c.ll

3–4 ↗

(On Diff #34338)

Just matching what is done in other tests - I'll strip them off.

[ab]

lib/Transforms/InstCombine/InstCombineCalls.cpp

846–857 ↗

(On Diff #34338)

Right: in case I wasn't clear, I'm not saying you should refactor the entire function for this patch, just to move the added ph2ps case block between storeu and cvtss2si.

llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp

case Intrinsic::ppc_qpx_qvstfd:
     if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
         32) {
       Type *OpPtrTy =
         PointerType::getUnqual(II->getArgOperand(0)->getType());
       Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
       return new StoreInst(II->getArgOperand(0), Ptr);
     }
     break;
+
   case Intrinsic::x86_sse_storeu_ps:
   case Intrinsic::x86_sse2_storeu_pd:
   case Intrinsic::x86_sse2_storeu_dq:
     // Turn X86 storeu -> store if the pointer is known aligned.
     if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
         16) {
       Type *OpPtrTy =
         PointerType::getUnqual(II->getArgOperand(1)->getType());
       Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
       return new StoreInst(II->getArgOperand(1), Ptr);
     }
     break;
 
+  case Intrinsic::x86_vcvtph2ps_128:
+  case Intrinsic::x86_vcvtph2ps_256: {
+    auto Arg = II->getArgOperand(0);
+    auto ArgType = cast<VectorType>(Arg->getType());
+    auto RetType = cast<VectorType>(II->getType());
+    unsigned ArgWidth = ArgType->getNumElements();
+    unsigned RetWidth = RetType->getNumElements();
+    assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
+    assert(ArgType->isIntOrIntVectorTy() &&
+           ArgType->getScalarSizeInBits() == 16 &&
+           "CVTPH2PS input type should be 16-bit integer vector");
+    assert(RetType->getScalarType()->isFloatTy() &&
+           "CVTPH2PS output type should be 32-bit float vector");
+
+    // Constant folding: Convert to generic half to single conversion.
+    if (auto CIZero = dyn_cast<ConstantAggregateZero>(Arg))
+      return ReplaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
+
+    if (auto CIHalf = dyn_cast<ConstantDataVector>(Arg)) {
+      auto VectorHalfAsShorts = Arg;
+      if (RetWidth < ArgWidth) {
+        SmallVector<int, 8> SubVecMask;
+        for (unsigned i = 0; i != RetWidth; ++i)
+          SubVecMask.push_back((int)i);
+        VectorHalfAsShorts = Builder->CreateShuffleVector(
+            Arg, UndefValue::get(ArgType), SubVecMask);
+      }
+
+      auto VectorHalfType =
+          VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
+      auto VectorHalfs =
+          Builder->CreateBitCast(VectorHalfAsShorts, VectorHalfType);
+      auto VectorFloats = Builder->CreateFPExt(VectorHalfs, RetType);
+      return ReplaceInstUsesWith(*II, VectorFloats);
+    }
+
+    // We only use the lowest lanes of the argument.
+    APInt DemandedElts = APInt::getLowBitsSet(ArgWidth, RetWidth);
+    APInt UndefElts(ArgWidth, 0);
+    if (Value *V = SimplifyDemandedVectorElts(Arg, DemandedElts, UndefElts)) {
+      II->setArgOperand(0, V);
+      return II;
+    }
+    break;
+  }
+
   case Intrinsic::x86_sse_cvtss2si:
   case Intrinsic::x86_sse_cvtss2si64:
   case Intrinsic::x86_sse_cvttss2si:
   case Intrinsic::x86_sse_cvttss2si64:
   case Intrinsic::x86_sse2_cvtsd2si:
   case Intrinsic::x86_sse2_cvtsd2si64:
   case Intrinsic::x86_sse2_cvttsd2si:
   case Intrinsic::x86_sse2_cvttsd2si64: {

llvm/trunk/test/Transforms/InstCombine/x86-f16c.ll

+; RUN: opt < %s -instcombine -S | FileCheck %s
+
+declare <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16>)
+declare <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16>)
+
+;
+; Vector Demanded Bits
+;
+
+; Only bottom 4 elements required.
+define <4 x float> @demand_vcvtph2ps_128(<8 x i16> %A) {
+; CHECK-LABEL: @demand_vcvtph2ps_128
+; CHECK-NEXT: %1 = tail call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %A)
+; CHECK-NEXT: ret <4 x float> %1
+  %1 = shufflevector <8 x i16> %A, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
+  %2 = tail call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %1)
+  ret <4 x float> %2
+}
+
+; All 8 elements required.
+define <8 x float> @demand_vcvtph2ps_256(<8 x i16> %A) {
+; CHECK-LABEL: @demand_vcvtph2ps_256
+; CHECK-NEXT: %1 = shufflevector <8 x i16> %A, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT: %2 = tail call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %1)
+; CHECK-NEXT: ret <8 x float> %2
+  %1 = shufflevector <8 x i16> %A, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
+  %2 = tail call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %1)
+  ret <8 x float> %2
+}
+
+;
+; Constant Folding
+;
+
+define <4 x float> @fold_vcvtph2ps_128() {
+; CHECK-LABEL: @fold_vcvtph2ps_128
+; CHECK-NEXT: ret <4 x float> <float 0.000000e+00, float 5.000000e-01, float 1.000000e+00, float -0.000000e+00>
+  %1 = tail call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> <i16 0, i16 14336, i16 15360, i16 32768, i16 16384, i16 31743, i16 48128, i16 49152>)
+  ret <4 x float> %1
+}
+
+define <8 x float> @fold_vcvtph2ps_256() {
+; CHECK-LABEL: @fold_vcvtph2ps_256
+; CHECK-NEXT: ret <8 x float> <float 0.000000e+00, float 5.000000e-01, float 1.000000e+00, float -0.000000e+00, float 2.000000e+00, float 6.550400e+04, float -1.000000e+00, float -2.000000e+00>
+  %1 = tail call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> <i16 0, i16 14336, i16 15360, i16 32768, i16 16384, i16 31743, i16 48128, i16 49152>)
+  ret <8 x float> %1
+}
+
+define <4 x float> @fold_vcvtph2ps_128_zero() {
+; CHECK-LABEL: @fold_vcvtph2ps_128_zero
+; CHECK-NEXT: ret <4 x float> zeroinitializer
+  %1 = tail call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
+  ret <4 x float> %1
+}
+
+define <8 x float> @fold_vcvtph2ps_256_zero() {
+; CHECK-LABEL: @fold_vcvtph2ps_256_zero
+; CHECK-NEXT: ret <8 x float> zeroinitializer
+  %1 = tail call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
+  ret <8 x float> %1
+}