Translate OpOrdered and OpUnordered to fcmp instructions

Instead of translating this instructions to OCL builtin calls, use core llvm fcmp instructions for better code preservation during translation. fixed tests separated fcmp test from OCL builtins
KhronosGroup · Oct 15, 2021 · 49e92c0 · 49e92c0
1 parent 098034e
commit 49e92c0
Show file tree

Hide file tree

Showing 4 changed files with 15 additions and 15 deletions.
diff --git a/lib/SPIRV/SPIRVReader.cpp b/lib/SPIRV/SPIRVReader.cpp
@@ -583,7 +583,7 @@ SPIRVToLLVM::transValue(const std::vector<SPIRVValue *> &BV, Function *F,
 
 bool SPIRVToLLVM::isSPIRVCmpInstTransToLLVMInst(SPIRVInstruction *BI) const {
   auto OC = BI->getOpCode();
-  return isCmpOpCode(OC) && !(OC >= OpLessOrGreater && OC <= OpUnordered);
+  return isCmpOpCode(OC) && OC != OpLessOrGreater;
 }
 
 void SPIRVToLLVM::setName(llvm::Value *V, SPIRVValue *BV) {

diff --git a/test/transcoding/fcmp.ll b/test/transcoding/fcmp.ll
@@ -229,9 +229,9 @@
 ; CHECK-LLVM: %r40 = fcmp oge float %a, %b
 ; CHECK-LLVM: %r41 = fcmp oge float %a, %b
 ; CHECK-LLVM: %r42 = fcmp oge float %a, %b
-; CHECK-LLVM: %r43 = call spir_func i32 @_Z9isorderedff(float %a, float %b)
-; CHECK-LLVM: %r44 = call spir_func i32 @_Z9isorderedff(float %a, float %b)
-; CHECK-LLVM: %r45 = call spir_func i32 @_Z9isorderedff(float %a, float %b)
+; CHECK-LLVM: %r43 = fcmp ord float %a, %b
+; CHECK-LLVM: %r44 = fcmp ord float %a, %b
+; CHECK-LLVM: %r45 = fcmp ord float %a, %b
 ; CHECK-LLVM: %r46 = fcmp ueq float %a, %b
 ; CHECK-LLVM: %r47 = fcmp ueq float %a, %b
 ; CHECK-LLVM: %r48 = fcmp ueq float %a, %b
@@ -274,9 +274,9 @@
 ; CHECK-LLVM: %r85 = fcmp uge float %a, %b
 ; CHECK-LLVM: %r86 = fcmp uge float %a, %b
 ; CHECK-LLVM: %r87 = fcmp uge float %a, %b
-; CHECK-LLVM: %r88 = call spir_func i32 @_Z11isunorderedff(float %a, float %b)
-; CHECK-LLVM: %r89 = call spir_func i32 @_Z11isunorderedff(float %a, float %b)
-; CHECK-LLVM: %r90 = call spir_func i32 @_Z11isunorderedff(float %a, float %b)
+; CHECK-LLVM: %r88 = fcmp uno float %a, %b
+; CHECK-LLVM: %r89 = fcmp uno float %a, %b
+; CHECK-LLVM: %r90 = fcmp uno float %a, %b
 
 target datalayout = "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
 target triple = "spir-unknown-unknown"

diff --git a/test/transcoding/relationals_float.ll b/test/transcoding/relationals_float.ll
@@ -12,16 +12,16 @@
 ; CHECK-LLVM: call spir_func i32 @_Z8isnormalf(
 ; CHECK-LLVM: call spir_func i32 @_Z7signbitf(
 ; CHECK-LLVM: call spir_func i32 @_Z13islessgreaterff(
-; CHECK-LLVM: call spir_func i32 @_Z9isorderedff(
-; CHECK-LLVM: call spir_func i32 @_Z11isunorderedff(
+; CHECK-LLVM: fcmp ord float
+; CHECK-LLVM: fcmp uno float
 
 ; CHECK-LLVM: call spir_func <2 x i32> @_Z8isfiniteDv2_f(
 ; CHECK-LLVM: call spir_func <2 x i32> @_Z5isnanDv2_f(
 ; CHECK-LLVM: call spir_func <2 x i32> @_Z5isinfDv2_f(
 ; CHECK-LLVM: call spir_func <2 x i32> @_Z8isnormalDv2_f(
 ; CHECK-LLVM: call spir_func <2 x i32> @_Z13islessgreaterDv2_fS_(
-; CHECK-LLVM: call spir_func <2 x i32> @_Z9isorderedDv2_fS_(
-; CHECK-LLVM: call spir_func <2 x i32> @_Z11isunorderedDv2_fS_(
+; CHECK-LLVM: fcmp ord <2 x float>
+; CHECK-LLVM: fcmp uno <2 x float>
 
 ; CHECK-SPIRV: 2 TypeBool [[BoolTypeID:[0-9]+]]
 ; CHECK-SPIRV: 4 TypeVector [[BoolVectorTypeID:[0-9]+]] [[BoolTypeID]] 2

diff --git a/test/transcoding/relationals_half.ll b/test/transcoding/relationals_half.ll
@@ -12,16 +12,16 @@
 ; CHECK-LLVM: call spir_func i32 @_Z8isnormalDh(
 ; CHECK-LLVM: call spir_func i32 @_Z7signbitDh(
 ; CHECK-LLVM: call spir_func i32 @_Z13islessgreaterDhDh(
-; CHECK-LLVM: call spir_func i32 @_Z9isorderedDhDh(
-; CHECK-LLVM: call spir_func i32 @_Z11isunorderedDhDh(
+; CHECK-LLVM: fcmp ord half
+; CHECK-LLVM: fcmp uno half
 
 ; CHECK-LLVM: call spir_func <2 x i16> @_Z8isfiniteDv2_Dh(
 ; CHECK-LLVM: call spir_func <2 x i16> @_Z5isnanDv2_Dh(
 ; CHECK-LLVM: call spir_func <2 x i16> @_Z5isinfDv2_Dh(
 ; CHECK-LLVM: call spir_func <2 x i16> @_Z8isnormalDv2_Dh(
 ; CHECK-LLVM: call spir_func <2 x i16> @_Z13islessgreaterDv2_DhS_(
-; CHECK-LLVM: call spir_func <2 x i16> @_Z9isorderedDv2_DhS_(
-; CHECK-LLVM: call spir_func <2 x i16> @_Z11isunorderedDv2_DhS_(
+; CHECK-LLVM: fcmp ord <2 x half>
+; CHECK-LLVM: fcmp uno <2 x half>
 
 ; CHECK-SPIRV: 2 TypeBool [[BoolTypeID:[0-9]+]]
 ; CHECK-SPIRV: 4 TypeVector [[BoolVectorTypeID:[0-9]+]] [[BoolTypeID]] 2