diff options
Diffstat (limited to 'test/CodeGen/X86/exedepsfix-broadcast.ll')
| -rw-r--r-- | test/CodeGen/X86/exedepsfix-broadcast.ll | 98 |
1 files changed, 54 insertions, 44 deletions
diff --git a/test/CodeGen/X86/exedepsfix-broadcast.ll b/test/CodeGen/X86/exedepsfix-broadcast.ll index ab92fe0d1d0c..992b3a395e7b 100644 --- a/test/CodeGen/X86/exedepsfix-broadcast.ll +++ b/test/CodeGen/X86/exedepsfix-broadcast.ll @@ -1,13 +1,16 @@ -; RUN: llc -O3 -mtriple=x86_64-apple-macosx -o - < %s -mattr=+avx2 -enable-unsafe-fp-math -mcpu=core2 | FileCheck %s +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py +; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx2 -enable-unsafe-fp-math | FileCheck %s + ; Check that the ExeDepsFix pass correctly fixes the domain for broadcast instructions. ; <rdar://problem/16354675> -; CHECK-LABEL: ExeDepsFix_broadcastss -; CHECK: broadcastss -; CHECK: vandps -; CHECK: vmaxps -; CHECK: ret define <4 x float> @ExeDepsFix_broadcastss(<4 x float> %arg, <4 x float> %arg2) { +; CHECK-LABEL: ExeDepsFix_broadcastss: +; CHECK: ## BB#0: +; CHECK-NEXT: vbroadcastss {{.*}}(%rip), %xmm2 +; CHECK-NEXT: vandps %xmm2, %xmm0, %xmm0 +; CHECK-NEXT: vmaxps %xmm1, %xmm0, %xmm0 +; CHECK-NEXT: retq %bitcast = bitcast <4 x float> %arg to <4 x i32> %and = and <4 x i32> %bitcast, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647> %floatcast = bitcast <4 x i32> %and to <4 x float> @@ -16,12 +19,13 @@ define <4 x float> @ExeDepsFix_broadcastss(<4 x float> %arg, <4 x float> %arg2) ret <4 x float> %max } -; CHECK-LABEL: ExeDepsFix_broadcastss256 -; CHECK: broadcastss -; CHECK: vandps -; CHECK: vmaxps -; CHECK: ret define <8 x float> @ExeDepsFix_broadcastss256(<8 x float> %arg, <8 x float> %arg2) { +; CHECK-LABEL: ExeDepsFix_broadcastss256: +; CHECK: ## BB#0: +; CHECK-NEXT: vbroadcastss {{.*}}(%rip), %ymm2 +; CHECK-NEXT: vandps %ymm2, %ymm0, %ymm0 +; CHECK-NEXT: vmaxps %ymm1, %ymm0, %ymm0 +; CHECK-NEXT: retq %bitcast = bitcast <8 x float> %arg to <8 x i32> %and = and <8 x i32> %bitcast, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647> %floatcast = bitcast <8 x i32> %and to <8 x float> @@ -30,13 +34,14 @@ define <8 x float> @ExeDepsFix_broadcastss256(<8 x float> %arg, <8 x float> %arg ret <8 x float> %max } - -; CHECK-LABEL: ExeDepsFix_broadcastss_inreg -; CHECK: broadcastss -; CHECK: vandps -; CHECK: vmaxps -; CHECK: ret define <4 x float> @ExeDepsFix_broadcastss_inreg(<4 x float> %arg, <4 x float> %arg2, i32 %broadcastvalue) { +; CHECK-LABEL: ExeDepsFix_broadcastss_inreg: +; CHECK: ## BB#0: +; CHECK-NEXT: vmovd %edi, %xmm2 +; CHECK-NEXT: vbroadcastss %xmm2, %xmm2 +; CHECK-NEXT: vandps %xmm2, %xmm0, %xmm0 +; CHECK-NEXT: vmaxps %xmm1, %xmm0, %xmm0 +; CHECK-NEXT: retq %bitcast = bitcast <4 x float> %arg to <4 x i32> %in = insertelement <4 x i32> undef, i32 %broadcastvalue, i32 0 %mask = shufflevector <4 x i32> %in, <4 x i32> undef, <4 x i32> zeroinitializer @@ -47,12 +52,14 @@ define <4 x float> @ExeDepsFix_broadcastss_inreg(<4 x float> %arg, <4 x float> % ret <4 x float> %max } -; CHECK-LABEL: ExeDepsFix_broadcastss256_inreg -; CHECK: broadcastss -; CHECK: vandps -; CHECK: vmaxps -; CHECK: ret define <8 x float> @ExeDepsFix_broadcastss256_inreg(<8 x float> %arg, <8 x float> %arg2, i32 %broadcastvalue) { +; CHECK-LABEL: ExeDepsFix_broadcastss256_inreg: +; CHECK: ## BB#0: +; CHECK-NEXT: vmovd %edi, %xmm2 +; CHECK-NEXT: vbroadcastss %xmm2, %ymm2 +; CHECK-NEXT: vandps %ymm2, %ymm0, %ymm0 +; CHECK-NEXT: vmaxps %ymm1, %ymm0, %ymm0 +; CHECK-NEXT: retq %bitcast = bitcast <8 x float> %arg to <8 x i32> %in = insertelement <8 x i32> undef, i32 %broadcastvalue, i32 0 %mask = shufflevector <8 x i32> %in, <8 x i32> undef, <8 x i32> zeroinitializer @@ -63,12 +70,13 @@ define <8 x float> @ExeDepsFix_broadcastss256_inreg(<8 x float> %arg, <8 x float ret <8 x float> %max } -; CHECK-LABEL: ExeDepsFix_broadcastsd ; In that case the broadcast is directly folded into vandpd. -; CHECK: vandpd -; CHECK: vmaxpd -; CHECK:ret define <2 x double> @ExeDepsFix_broadcastsd(<2 x double> %arg, <2 x double> %arg2) { +; CHECK-LABEL: ExeDepsFix_broadcastsd: +; CHECK: ## BB#0: +; CHECK-NEXT: vandpd {{.*}}(%rip), %xmm0, %xmm0 +; CHECK-NEXT: vmaxpd %xmm1, %xmm0, %xmm0 +; CHECK-NEXT: retq %bitcast = bitcast <2 x double> %arg to <2 x i64> %and = and <2 x i64> %bitcast, <i64 2147483647, i64 2147483647> %floatcast = bitcast <2 x i64> %and to <2 x double> @@ -77,12 +85,13 @@ define <2 x double> @ExeDepsFix_broadcastsd(<2 x double> %arg, <2 x double> %arg ret <2 x double> %max } -; CHECK-LABEL: ExeDepsFix_broadcastsd256 -; CHECK: broadcastsd -; CHECK: vandpd -; CHECK: vmaxpd -; CHECK: ret define <4 x double> @ExeDepsFix_broadcastsd256(<4 x double> %arg, <4 x double> %arg2) { +; CHECK-LABEL: ExeDepsFix_broadcastsd256: +; CHECK: ## BB#0: +; CHECK-NEXT: vbroadcastsd {{.*}}(%rip), %ymm2 +; CHECK-NEXT: vandpd %ymm2, %ymm0, %ymm0 +; CHECK-NEXT: vmaxpd %ymm1, %ymm0, %ymm0 +; CHECK-NEXT: retq %bitcast = bitcast <4 x double> %arg to <4 x i64> %and = and <4 x i64> %bitcast, <i64 2147483647, i64 2147483647, i64 2147483647, i64 2147483647> %floatcast = bitcast <4 x i64> %and to <4 x double> @@ -91,16 +100,16 @@ define <4 x double> @ExeDepsFix_broadcastsd256(<4 x double> %arg, <4 x double> % ret <4 x double> %max } - -; CHECK-LABEL: ExeDepsFix_broadcastsd_inreg ; ExeDepsFix works top down, thus it coalesces vpunpcklqdq domain with ; vpand and there is nothing more you can do to match vmaxpd. -; CHECK: vmovq -; CHECK: vpbroadcastq -; CHECK: vpand -; CHECK: vmaxpd -; CHECK: ret define <2 x double> @ExeDepsFix_broadcastsd_inreg(<2 x double> %arg, <2 x double> %arg2, i64 %broadcastvalue) { +; CHECK-LABEL: ExeDepsFix_broadcastsd_inreg: +; CHECK: ## BB#0: +; CHECK-NEXT: vmovq %rdi, %xmm2 +; CHECK-NEXT: vpbroadcastq %xmm2, %xmm2 +; CHECK-NEXT: vpand %xmm2, %xmm0, %xmm0 +; CHECK-NEXT: vmaxpd %xmm1, %xmm0, %xmm0 +; CHECK-NEXT: retq %bitcast = bitcast <2 x double> %arg to <2 x i64> %in = insertelement <2 x i64> undef, i64 %broadcastvalue, i32 0 %mask = shufflevector <2 x i64> %in, <2 x i64> undef, <2 x i32> zeroinitializer @@ -111,12 +120,14 @@ define <2 x double> @ExeDepsFix_broadcastsd_inreg(<2 x double> %arg, <2 x double ret <2 x double> %max } -; CHECK-LABEL: ExeDepsFix_broadcastsd256_inreg -; CHECK: broadcastsd -; CHECK: vandpd -; CHECK: vmaxpd -; CHECK: ret define <4 x double> @ExeDepsFix_broadcastsd256_inreg(<4 x double> %arg, <4 x double> %arg2, i64 %broadcastvalue) { +; CHECK-LABEL: ExeDepsFix_broadcastsd256_inreg: +; CHECK: ## BB#0: +; CHECK-NEXT: vmovq %rdi, %xmm2 +; CHECK-NEXT: vbroadcastsd %xmm2, %ymm2 +; CHECK-NEXT: vandpd %ymm2, %ymm0, %ymm0 +; CHECK-NEXT: vmaxpd %ymm1, %ymm0, %ymm0 +; CHECK-NEXT: retq %bitcast = bitcast <4 x double> %arg to <4 x i64> %in = insertelement <4 x i64> undef, i64 %broadcastvalue, i32 0 %mask = shufflevector <4 x i64> %in, <4 x i64> undef, <4 x i32> zeroinitializer @@ -126,4 +137,3 @@ define <4 x double> @ExeDepsFix_broadcastsd256_inreg(<4 x double> %arg, <4 x dou %max = select <4 x i1> %max_is_x, <4 x double> %floatcast, <4 x double> %arg2 ret <4 x double> %max } - |