[X86] Add broadcast load unfolding support for VMAXPS/PD and VMINPS/PD.

llvm-svn: 371363

diff --git a/llvm/lib/Target/X86/X86InstrFoldTables.cpp b/llvm/lib/Target/X86/X86InstrFoldTables.cpp
index 63474f9..bddb8b8 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrFoldTables.cpp
+++ b/llvm/lib/Target/X86/X86InstrFoldTables.cpp
@@ -5258,6 +5258,30 @@ static const X86MemoryFoldTableEntry BroadcastFoldTable2[] = {
   { X86::VDIVPSZ128rr,   X86::VDIVPSZ128rmb,   TB_BCAST_SS },
   { X86::VDIVPSZ256rr,   X86::VDIVPSZ256rmb,   TB_BCAST_SS },
   { X86::VDIVPSZrr,      X86::VDIVPSZrmb,      TB_BCAST_SS },
+  { X86::VMAXCPDZ128rr,  X86::VMAXCPDZ128rmb,  TB_BCAST_SD },
+  { X86::VMAXCPDZ256rr,  X86::VMAXCPDZ256rmb,  TB_BCAST_SD },
+  { X86::VMAXCPDZrr,     X86::VMAXCPDZrmb,     TB_BCAST_SD },
+  { X86::VMAXCPSZ128rr,  X86::VMAXCPSZ128rmb,  TB_BCAST_SS },
+  { X86::VMAXCPSZ256rr,  X86::VMAXCPSZ256rmb,  TB_BCAST_SS },
+  { X86::VMAXCPSZrr,     X86::VMAXCPSZrmb,     TB_BCAST_SS },
+  { X86::VMAXPDZ128rr,   X86::VMAXPDZ128rmb,   TB_BCAST_SD },
+  { X86::VMAXPDZ256rr,   X86::VMAXPDZ256rmb,   TB_BCAST_SD },
+  { X86::VMAXPDZrr,      X86::VMAXPDZrmb,      TB_BCAST_SD },
+  { X86::VMAXPSZ128rr,   X86::VMAXPSZ128rmb,   TB_BCAST_SS },
+  { X86::VMAXPSZ256rr,   X86::VMAXPSZ256rmb,   TB_BCAST_SS },
+  { X86::VMAXPSZrr,      X86::VMAXPSZrmb,      TB_BCAST_SS },
+  { X86::VMINCPDZ128rr,  X86::VMINCPDZ128rmb,  TB_BCAST_SD },
+  { X86::VMINCPDZ256rr,  X86::VMINCPDZ256rmb,  TB_BCAST_SD },
+  { X86::VMINCPDZrr,     X86::VMINCPDZrmb,     TB_BCAST_SD },
+  { X86::VMINCPSZ128rr,  X86::VMINCPSZ128rmb,  TB_BCAST_SS },
+  { X86::VMINCPSZ256rr,  X86::VMINCPSZ256rmb,  TB_BCAST_SS },
+  { X86::VMINCPSZrr,     X86::VMINCPSZrmb,     TB_BCAST_SS },
+  { X86::VMINPDZ128rr,   X86::VMINPDZ128rmb,   TB_BCAST_SD },
+  { X86::VMINPDZ256rr,   X86::VMINPDZ256rmb,   TB_BCAST_SD },
+  { X86::VMINPDZrr,      X86::VMINPDZrmb,      TB_BCAST_SD },
+  { X86::VMINPSZ128rr,   X86::VMINPSZ128rmb,   TB_BCAST_SS },
+  { X86::VMINPSZ256rr,   X86::VMINPSZ256rmb,   TB_BCAST_SS },
+  { X86::VMINPSZrr,      X86::VMINPSZrmb,      TB_BCAST_SS },
   { X86::VMULPDZ128rr,   X86::VMULPDZ128rmb,   TB_BCAST_SD },
   { X86::VMULPDZ256rr,   X86::VMULPDZ256rmb,   TB_BCAST_SD },
   { X86::VMULPDZrr,      X86::VMULPDZrmb,      TB_BCAST_SD },

diff --git a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
index e1373e0..569577f 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
+++ b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
@@ -2079,12 +2079,13 @@ define void @bcast_unfold_fmax_v4f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmax_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB60_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vmaxps {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovups %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm1
+; CHECK-NEXT:    vmaxps %xmm0, %xmm1, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB60_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2113,12 +2114,13 @@ define void @bcast_unfold_fmax_v8f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmax_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; 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:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB61_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vmaxps {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovups %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm1
+; CHECK-NEXT:    vmaxps %ymm0, %ymm1, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB61_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2148,12 +2150,13 @@ define void @bcast_unfold_fmax_v16f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmax_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; 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:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB62_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vmaxps {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovups %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm1
+; CHECK-NEXT:    vmaxps %zmm0, %zmm1, %zmm1
+; CHECK-NEXT:    vmovups %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB62_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2218,12 +2221,13 @@ define void @bcast_unfold_fmax_v4f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmax_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB64_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vmaxpd {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovupd %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm1
+; CHECK-NEXT:    vmaxpd %ymm0, %ymm1, %ymm1
+; CHECK-NEXT:    vmovupd %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB64_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2253,12 +2257,13 @@ define void @bcast_unfold_fmax_v8f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmax_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; 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:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB65_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vmaxpd {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovupd %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm1
+; CHECK-NEXT:    vmaxpd %zmm0, %zmm1, %zmm1
+; CHECK-NEXT:    vmovupd %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB65_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2288,12 +2293,13 @@ define void @bcast_unfold_fmin_v4f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmin_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB66_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vminps {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovups %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm1
+; CHECK-NEXT:    vminps %xmm0, %xmm1, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB66_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2322,12 +2328,13 @@ define void @bcast_unfold_fmin_v8f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmin_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; 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:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB67_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vminps {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovups %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm1
+; CHECK-NEXT:    vminps %ymm0, %ymm1, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB67_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2357,12 +2364,13 @@ define void @bcast_unfold_fmin_v16f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmin_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; 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:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB68_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vminps {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovups %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm1
+; CHECK-NEXT:    vminps %zmm0, %zmm1, %zmm1
+; CHECK-NEXT:    vmovups %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB68_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2427,12 +2435,13 @@ define void @bcast_unfold_fmin_v4f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmin_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB70_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vminpd {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovupd %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm1
+; CHECK-NEXT:    vminpd %ymm0, %ymm1, %ymm1
+; CHECK-NEXT:    vmovupd %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB70_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
@@ -2462,12 +2471,13 @@ define void @bcast_unfold_fmin_v8f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fmin_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; 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:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB71_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vminpd {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovupd %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm1
+; CHECK-NEXT:    vminpd %zmm0, %zmm1, %zmm1
+; CHECK-NEXT:    vmovupd %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB71_1
 ; CHECK-NEXT:  # %bb.2: # %bb10