diff options
Diffstat (limited to 'test/CodeGen/ARM/vfcmp.ll')
-rw-r--r-- | test/CodeGen/ARM/vfcmp.ll | 101 |
1 files changed, 72 insertions, 29 deletions
diff --git a/test/CodeGen/ARM/vfcmp.ll b/test/CodeGen/ARM/vfcmp.ll index 58c2068..6946d02 100644 --- a/test/CodeGen/ARM/vfcmp.ll +++ b/test/CodeGen/ARM/vfcmp.ll @@ -1,96 +1,139 @@ -; RUN: llvm-as < %s | llc -march=arm -mattr=+neon > %t -; RUN: grep {vceq\\.f32} %t | count 1 -; RUN: grep {vcgt\\.f32} %t | count 9 -; RUN: grep {vcge\\.f32} %t | count 5 -; RUN: grep vorr %t | count 4 -; RUN: grep vmvn %t | count 7 +; RUN: llc < %s -march=arm -mattr=+neon | FileCheck %s -; This tests vfcmp operations that do not map directly to NEON instructions. +; This tests fcmp operations that do not map directly to NEON instructions. ; une is implemented with VCEQ/VMVN define <2 x i32> @vcunef32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcunef32: +;CHECK: vceq.f32 +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp une <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp une <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; olt is implemented with VCGT define <2 x i32> @vcoltf32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcoltf32: +;CHECK: vcgt.f32 %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp olt <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp olt <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; ole is implemented with VCGE define <2 x i32> @vcolef32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcolef32: +;CHECK: vcge.f32 %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp ole <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp ole <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; uge is implemented with VCGT/VMVN define <2 x i32> @vcugef32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcugef32: +;CHECK: vcgt.f32 +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp uge <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp uge <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; ule is implemented with VCGT/VMVN define <2 x i32> @vculef32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vculef32: +;CHECK: vcgt.f32 +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp ule <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp ule <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; ugt is implemented with VCGE/VMVN define <2 x i32> @vcugtf32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcugtf32: +;CHECK: vcge.f32 +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp ugt <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp ugt <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; ult is implemented with VCGE/VMVN define <2 x i32> @vcultf32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcultf32: +;CHECK: vcge.f32 +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp ult <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp ult <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; ueq is implemented with VCGT/VCGT/VORR/VMVN define <2 x i32> @vcueqf32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcueqf32: +;CHECK: vcgt.f32 +;CHECK-NEXT: vcgt.f32 +;CHECK-NEXT: vorr +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp ueq <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp ueq <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; one is implemented with VCGT/VCGT/VORR define <2 x i32> @vconef32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vconef32: +;CHECK: vcgt.f32 +;CHECK-NEXT: vcgt.f32 +;CHECK-NEXT: vorr %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp one <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp one <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; uno is implemented with VCGT/VCGE/VORR/VMVN define <2 x i32> @vcunof32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcunof32: +;CHECK: vcge.f32 +;CHECK-NEXT: vcgt.f32 +;CHECK-NEXT: vorr +;CHECK-NEXT: vmvn %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp uno <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp uno <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } ; ord is implemented with VCGT/VCGE/VORR define <2 x i32> @vcordf32(<2 x float>* %A, <2 x float>* %B) nounwind { +;CHECK: vcordf32: +;CHECK: vcge.f32 +;CHECK-NEXT: vcgt.f32 +;CHECK-NEXT: vorr %tmp1 = load <2 x float>* %A %tmp2 = load <2 x float>* %B - %tmp3 = vfcmp ord <2 x float> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %tmp3 = fcmp ord <2 x float> %tmp1, %tmp2 + %tmp4 = sext <2 x i1> %tmp3 to <2 x i32> + ret <2 x i32> %tmp4 } |