From 4a0dc124f30cf0becf82c84999e2ae18827605e0 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Wed, 29 Nov 2017 00:32:09 +0000
Subject: [PATCH] [X86] Promote fp_to_sint v16f32->v16i16/v16i8 to avoid
 scalarization.

llvm-svn: 319266
---
 lib/Target/X86/X86ISelLowering.cpp    |  2 +
 test/Analysis/CostModel/X86/fptosi.ll |  4 +-
 test/CodeGen/X86/avx512-cvt.ll        | 99 ++-------------------------
 3 files changed, 8 insertions(+), 97 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index bdc312260bc..d15c2825154 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1173,6 +1173,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     }
 
     setOperationAction(ISD::FP_TO_SINT,         MVT::v16i32, Legal);
+    setOperationAction(ISD::FP_TO_SINT,         MVT::v16i16, Promote);
+    setOperationAction(ISD::FP_TO_SINT,         MVT::v16i8, Promote);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v16i32, Legal);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v16i8, Promote);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v16i16, Promote);
diff --git a/test/Analysis/CostModel/X86/fptosi.ll b/test/Analysis/CostModel/X86/fptosi.ll
index d5e21f8685a..dc0a041f94c 100644
--- a/test/Analysis/CostModel/X86/fptosi.ll
+++ b/test/Analysis/CostModel/X86/fptosi.ll
@@ -224,7 +224,7 @@ define i32 @fptosi_float_i16(i32 %arg) {
   ; SSE42: cost of 7 {{.*}} %V16I16 = fptosi
   ; AVX1: cost of 3 {{.*}} %V16I16 = fptosi
   ; AVX2: cost of 3 {{.*}} %V16I16 = fptosi
-  ; AVX512: cost of 48 {{.*}} %V16I16 = fptosi
+  ; AVX512: cost of 1 {{.*}} %V16I16 = fptosi
   %V16I16 = fptosi <16 x float> undef to <16 x i16>
 
   ret i32 undef
@@ -254,7 +254,7 @@ define i32 @fptosi_float_i8(i32 %arg) {
   ; SSE42: cost of 7 {{.*}} %V16I8 = fptosi
   ; AVX1: cost of 15 {{.*}} %V16I8 = fptosi
   ; AVX2: cost of 15 {{.*}} %V16I8 = fptosi
-  ; AVX512: cost of 48 {{.*}} %V16I8 = fptosi
+  ; AVX512: cost of 1 {{.*}} %V16I8 = fptosi
   %V16I8 = fptosi <16 x float> undef to <16 x i8>
 
   ret i32 undef
diff --git a/test/CodeGen/X86/avx512-cvt.ll b/test/CodeGen/X86/avx512-cvt.ll
index afb94614f56..9c17ea603ac 100644
--- a/test/CodeGen/X86/avx512-cvt.ll
+++ b/test/CodeGen/X86/avx512-cvt.ll
@@ -433,53 +433,8 @@ define <16 x i32> @f64to16si(<16 x float> %a) nounwind {
 define <16 x i8> @f32to16sc(<16 x float> %f) {
 ; ALL-LABEL: f32to16sc:
 ; ALL:       # BB#0:
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm1, %eax
-; ALL-NEXT:    vcvttss2si %xmm0, %ecx
-; ALL-NEXT:    vmovd %ecx, %xmm1
-; ALL-NEXT:    vpinsrb $1, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $2, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilps {{.*#+}} xmm2 = xmm0[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $3, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrb $5, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm3 = xmm2[1,0]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrb $6, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $7, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vextractf32x4 $2, %zmm0, %xmm2
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrb $9, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm3 = xmm2[1,0]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrb $10, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $11, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vextractf32x4 $3, %zmm0, %xmm0
-; ALL-NEXT:    vcvttss2si %xmm0, %eax
-; ALL-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $13, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrb $14, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm0, %eax
-; ALL-NEXT:    vpinsrb $15, %eax, %xmm1, %xmm0
+; ALL-NEXT:    vcvttps2dq %zmm0, %zmm0
+; ALL-NEXT:    vpmovdb %zmm0, %xmm0
 ; ALL-NEXT:    vzeroupper
 ; ALL-NEXT:    retq
   %res = fptosi <16 x float> %f to <16 x i8>
@@ -489,54 +444,8 @@ define <16 x i8> @f32to16sc(<16 x float> %f) {
 define <16 x i16> @f32to16ss(<16 x float> %f) {
 ; ALL-LABEL: f32to16ss:
 ; ALL:       # BB#0:
-; ALL-NEXT:    vextractf32x4 $2, %zmm0, %xmm1
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vcvttss2si %xmm1, %ecx
-; ALL-NEXT:    vmovd %ecx, %xmm2
-; ALL-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm3 = xmm1[1,0]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm1, %eax
-; ALL-NEXT:    vpinsrw $3, %eax, %xmm2, %xmm1
-; ALL-NEXT:    vextractf32x4 $3, %zmm0, %xmm2
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrw $4, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $5, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm3 = xmm2[1,0]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $6, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vpinsrw $7, %eax, %xmm1, %xmm1
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm2, %eax
-; ALL-NEXT:    vcvttss2si %xmm0, %ecx
-; ALL-NEXT:    vmovd %ecx, %xmm2
-; ALL-NEXT:    vpinsrw $1, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm3 = xmm0[1,0]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $2, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vpermilps {{.*#+}} xmm3 = xmm0[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $3, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; ALL-NEXT:    vcvttss2si %xmm0, %eax
-; ALL-NEXT:    vpinsrw $4, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm0[1,1,3,3]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $5, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vpermilpd {{.*#+}} xmm3 = xmm0[1,0]
-; ALL-NEXT:    vcvttss2si %xmm3, %eax
-; ALL-NEXT:    vpinsrw $6, %eax, %xmm2, %xmm2
-; ALL-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; ALL-NEXT:    vcvttss2si %xmm0, %eax
-; ALL-NEXT:    vpinsrw $7, %eax, %xmm2, %xmm0
-; ALL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
+; ALL-NEXT:    vcvttps2dq %zmm0, %zmm0
+; ALL-NEXT:    vpmovdw %zmm0, %ymm0
 ; ALL-NEXT:    retq
   %res = fptosi <16 x float> %f to <16 x i16>
   ret <16 x i16> %res