[llvm] 86b5e25 - [SelectionDAG][X86] Support f16 in getReciprocalOpName.

[Craig Topper via llvm-commits]

Author: Craig Topper
Date: 2022-02-18T15:36:35-08:00
New Revision: 86b5e256628ae49193ad9962626a73bafeda2883

URL: https://github.com/llvm/llvm-project/commit/86b5e256628ae49193ad9962626a73bafeda2883
DIFF: https://github.com/llvm/llvm-project/commit/86b5e256628ae49193ad9962626a73bafeda2883.diff

LOG: [SelectionDAG][X86] Support f16 in getReciprocalOpName.

If the "reciprocal-estimates" attribute is present and it doesn't
contain "all", "none", or "default", we previously crashed on f16
operations.

This patch addes an 'h' suffix' to prevent the crash.

I've added simple tests that just enable the estimate for all
vec-sqrt and one test case that explicitly tests the new 'h' suffix
to override the default steps.

There may be some frontend change needed to, but I haven't checked
that yet.

Differential Revision: https://reviews.llvm.org/D120158

Added: 
    

Modified: 
    llvm/lib/CodeGen/TargetLoweringBase.cpp
    llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index 3a7e82c9038c1..700c11a66904f 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -2072,9 +2072,11 @@ static std::string getReciprocalOpName(bool IsSqrt, EVT VT) {
 
   Name += IsSqrt ? "sqrt" : "div";
 
-  // TODO: Handle "half" or other float types?
+  // TODO: Handle other float types?
   if (VT.getScalarType() == MVT::f64) {
     Name += "d";
+  } else if (VT.getScalarType() == MVT::f16) {
+    Name += "h";
   } else {
     assert(VT.getScalarType() == MVT::f32 &&
            "Unexpected FP type for reciprocal estimate");

diff  --git a/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll b/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll
index 6a5c3e243209a..61b483329a4d9 100644
--- a/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll
@@ -35,6 +35,32 @@ define <32 x half> @test_sqrt_ph_512_fast(<32 x half> %a0, <32 x half> %a1) {
   ret <32 x half> %2
 }
 
+define <32 x half> @test_sqrt_ph_512_fast_estimate_attribute(<32 x half> %a0, <32 x half> %a1) "reciprocal-estimates"="vec-sqrt" {
+; CHECK-LABEL: test_sqrt_ph_512_fast_estimate_attribute:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vrsqrtph %zmm0, %zmm0
+; CHECK-NEXT:    vmulph %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    retq
+  %1 = call fast <32 x half> @llvm.sqrt.v32f16(<32 x half> %a0)
+  %2 = fdiv fast <32 x half> %a1, %1
+  ret <32 x half> %2
+}
+
+define <32 x half> @test_sqrt_ph_512_fast_estimate_attribute_2(<32 x half> %a0, <32 x half> %a1) "reciprocal-estimates"="vec-sqrth:1" {
+; CHECK-LABEL: test_sqrt_ph_512_fast_estimate_attribute_2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vrsqrtph %zmm0, %zmm2
+; CHECK-NEXT:    vmulph %zmm2, %zmm0, %zmm0
+; CHECK-NEXT:    vfmadd213ph {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to32}, %zmm2, %zmm0
+; CHECK-NEXT:    vmulph {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to32}, %zmm2, %zmm2
+; CHECK-NEXT:    vmulph %zmm1, %zmm0, %zmm0
+; CHECK-NEXT:    vmulph %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %1 = call fast <32 x half> @llvm.sqrt.v32f16(<32 x half> %a0)
+  %2 = fdiv fast <32 x half> %a1, %1
+  ret <32 x half> %2
+}
+
 define <32 x half> @test_mask_sqrt_ph_512(<32 x half> %a0, <32 x half> %passthru, i32 %mask) {
 ; CHECK-LABEL: test_mask_sqrt_ph_512:
 ; CHECK:       # %bb.0: