From 0e533ca4bb82fee0b1c25853129a2f5f80dbf97b Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Tue, 10 Sep 2019 05:49:53 +0000
Subject: [PATCH] [X86] Add broadcast load unfolding support for VCMPPS/PD.

llvm-svn: 371487
---
 llvm/lib/Target/X86/X86InstrFoldTables.cpp       |  6 +++
 llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll | 55 +++++++++++++-----------
 2 files changed, 36 insertions(+), 25 deletions(-)

diff --git a/llvm/lib/Target/X86/X86InstrFoldTables.cpp b/llvm/lib/Target/X86/X86InstrFoldTables.cpp
index 07bbf4c..0f89a80 100644
--- a/llvm/lib/Target/X86/X86InstrFoldTables.cpp
+++ b/llvm/lib/Target/X86/X86InstrFoldTables.cpp
@@ -5252,6 +5252,12 @@ static const X86MemoryFoldTableEntry BroadcastFoldTable2[] = {
   { X86::VADDPSZ128rr,   X86::VADDPSZ128rmb,   TB_BCAST_SS },
   { X86::VADDPSZ256rr,   X86::VADDPSZ256rmb,   TB_BCAST_SS },
   { X86::VADDPSZrr,      X86::VADDPSZrmb,      TB_BCAST_SS },
+  { X86::VCMPPDZ128rri,  X86::VCMPPDZ128rmbi,  TB_BCAST_SD },
+  { X86::VCMPPDZ256rri,  X86::VCMPPDZ256rmbi,  TB_BCAST_SD },
+  { X86::VCMPPDZrri,     X86::VCMPPDZrmbi,     TB_BCAST_SD },
+  { X86::VCMPPSZ128rri,  X86::VCMPPSZ128rmbi,  TB_BCAST_SS },
+  { X86::VCMPPSZ256rri,  X86::VCMPPSZ256rmbi,  TB_BCAST_SS },
+  { X86::VCMPPSZrri,     X86::VCMPPSZrmbi,     TB_BCAST_SS },
   { X86::VDIVPDZ128rr,   X86::VDIVPDZ128rmb,   TB_BCAST_SD },
   { X86::VDIVPDZ256rr,   X86::VDIVPDZ256rmb,   TB_BCAST_SD },
   { X86::VDIVPDZrr,      X86::VDIVPDZrmb,      TB_BCAST_SD },
diff --git a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
index a798102..fb097dd 100644
--- a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
+++ b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
@@ -4227,14 +4227,15 @@ define void @bcast_unfold_cmp_v4f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_cmp_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
-; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0]
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB120_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm1
-; CHECK-NEXT:    vcmpltps {{.*}}(%rip){1to4}, %xmm1, %k1
-; CHECK-NEXT:    vblendmps %xmm1, %xmm0, %xmm1 {%k1}
-; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm2
+; CHECK-NEXT:    vcmpltps %xmm0, %xmm2, %k1
+; CHECK-NEXT:    vblendmps %xmm2, %xmm1, %xmm2 {%k1}
+; CHECK-NEXT:    vmovups %xmm2, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB120_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -4263,14 +4264,15 @@ define void @bcast_unfold_cmp_v8f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_cmp_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
-; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB121_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm1
-; CHECK-NEXT:    vcmpltps {{.*}}(%rip){1to8}, %ymm1, %k1
-; CHECK-NEXT:    vblendmps %ymm1, %ymm0, %ymm1 {%k1}
-; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm2
+; CHECK-NEXT:    vcmpltps %ymm0, %ymm2, %k1
+; CHECK-NEXT:    vblendmps %ymm2, %ymm1, %ymm2 {%k1}
+; CHECK-NEXT:    vmovups %ymm2, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB121_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -4300,14 +4302,15 @@ define void @bcast_unfold_cmp_v16f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_cmp_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
-; CHECK-NEXT:    vbroadcastss {{.*#+}} zmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
+; CHECK-NEXT:    vbroadcastss {{.*#+}} zmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
+; CHECK-NEXT:    vbroadcastss {{.*#+}} zmm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB122_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm1
-; CHECK-NEXT:    vcmpltps {{.*}}(%rip){1to16}, %zmm1, %k1
-; CHECK-NEXT:    vblendmps %zmm1, %zmm0, %zmm1 {%k1}
-; CHECK-NEXT:    vmovups %zmm1, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm2
+; CHECK-NEXT:    vcmpltps %zmm0, %zmm2, %k1
+; CHECK-NEXT:    vblendmps %zmm2, %zmm1, %zmm2 {%k1}
+; CHECK-NEXT:    vmovups %zmm2, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB122_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -4374,14 +4377,15 @@ define void @bcast_unfold_cmp_v4f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_cmp_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
-; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0]
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB124_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm1
-; CHECK-NEXT:    vcmpltpd {{.*}}(%rip){1to4}, %ymm1, %k1
-; CHECK-NEXT:    vblendmpd %ymm1, %ymm0, %ymm1 {%k1}
-; CHECK-NEXT:    vmovupd %ymm1, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm2
+; CHECK-NEXT:    vcmpltpd %ymm0, %ymm2, %k1
+; CHECK-NEXT:    vblendmpd %ymm2, %ymm1, %ymm2 {%k1}
+; CHECK-NEXT:    vmovupd %ymm2, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB124_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -4411,14 +4415,15 @@ define void @bcast_unfold_cmp_v8f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_cmp_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
-; CHECK-NEXT:    vbroadcastsd {{.*#+}} zmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} zmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} zmm1 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB125_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm1
-; CHECK-NEXT:    vcmpltpd {{.*}}(%rip){1to8}, %zmm1, %k1
-; CHECK-NEXT:    vblendmpd %zmm1, %zmm0, %zmm1 {%k1}
-; CHECK-NEXT:    vmovupd %zmm1, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm2
+; CHECK-NEXT:    vcmpltpd %zmm0, %zmm2, %k1
+; CHECK-NEXT:    vblendmpd %zmm2, %zmm1, %zmm2 {%k1}
+; CHECK-NEXT:    vmovupd %zmm2, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB125_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
-- 
2.7.4