[llvm] r268375 - [AVX512] Add support for commutative MAX/MIN . In general VMAX{PS, PD} and VMIN{PS, PD} instruction are not commutative . In combine pass only if UnsafeFPMath are used VMAX/VMAX are converted to commutative nodes VMAXC/VMAXC.

[Igor Breger via llvm-commits]

Author: ibreger
Date: Tue May  3 06:51:45 2016
New Revision: 268375

URL: http://llvm.org/viewvc/llvm-project?rev=268375&view=rev
Log:
[AVX512] Add support for commutative MAX/MIN . In general VMAX{PS,PD} and VMIN{PS,PD} instruction are not commutative . In combine pass only if UnsafeFPMath are used VMAX/VMAX are converted to commutative nodes VMAXC/VMAXC.

Differential Revision: http://reviews.llvm.org/D19860

Added:
    llvm/trunk/test/CodeGen/X86/avx512-unsafe-fp-math.ll
Modified:
    llvm/trunk/lib/Target/X86/X86InstrAVX512.td

Modified: llvm/trunk/lib/Target/X86/X86InstrAVX512.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86InstrAVX512.td?rev=268375&r1=268374&r2=268375&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86InstrAVX512.td (original)
+++ llvm/trunk/lib/Target/X86/X86InstrAVX512.td Tue May  3 06:51:45 2016
@@ -3742,10 +3742,14 @@ defm VSUB : avx512_fp_binop_p<0x5C, "vsu
             avx512_fp_binop_p_round<0x5C, "vsub", X86fsubRnd>;
 defm VDIV : avx512_fp_binop_p<0x5E, "vdiv", fdiv, HasAVX512>,
             avx512_fp_binop_p_round<0x5E, "vdiv", X86fdivRnd>;
-defm VMIN : avx512_fp_binop_p<0x5D, "vmin", X86fmin, HasAVX512, 1>,
+defm VMIN : avx512_fp_binop_p<0x5D, "vmin", X86fmin, HasAVX512, 0>,
             avx512_fp_binop_p_sae<0x5D, "vmin", X86fminRnd>;
-defm VMAX : avx512_fp_binop_p<0x5F, "vmax", X86fmax, HasAVX512, 1>,
+defm VMAX : avx512_fp_binop_p<0x5F, "vmax", X86fmax, HasAVX512, 0>,
             avx512_fp_binop_p_sae<0x5F, "vmax", X86fmaxRnd>;
+let isCodeGenOnly = 1 in {
+  defm VMINC : avx512_fp_binop_p<0x5D, "vmin", X86fminc, HasAVX512, 1>;
+  defm VMAXC : avx512_fp_binop_p<0x5F, "vmax", X86fmaxc, HasAVX512, 1>;
+}
 defm VAND  : avx512_fp_binop_p<0x54, "vand", X86fand, HasDQI, 1>;
 defm VANDN : avx512_fp_binop_p<0x55, "vandn", X86fandn, HasDQI, 0>;
 defm VOR   : avx512_fp_binop_p<0x56, "vor", X86for, HasDQI, 1>;

Added: llvm/trunk/test/CodeGen/X86/avx512-unsafe-fp-math.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx512-unsafe-fp-math.ll?rev=268375&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avx512-unsafe-fp-math.ll (added)
+++ llvm/trunk/test/CodeGen/X86/avx512-unsafe-fp-math.ll Tue May  3 06:51:45 2016
@@ -0,0 +1,71 @@
+; NOTE: Assertions have been autogenerated by update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math -mattr=+avx512f | FileCheck %s --check-prefix=CHECK_UNSAFE --check-prefix=AVX512F_UNSAFE
+; RUN: llc < %s -mtriple=x86_64 -mattr=+avx512f | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512
+
+define <16 x float> @test_max_v16f32(<16 x float> * %a_ptr, <16 x float> %b)  {
+; CHECK_UNSAFE-LABEL: test_max_v16f32:
+; CHECK_UNSAFE:       # BB#0:
+; CHECK_UNSAFE-NEXT:    vmaxps (%rdi), %zmm0, %zmm0
+; CHECK_UNSAFE-NEXT:    retq
+;
+; CHECK-LABEL: test_max_v16f32:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vmovaps (%rdi), %zmm1
+; CHECK-NEXT:    vmaxps %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    retq
+  %a = load <16 x float>, <16 x float>* %a_ptr
+  %tmp = fcmp fast ogt <16 x float> %a, %b
+  %tmp4 = select <16 x i1> %tmp, <16 x float> %a, <16 x float> %b
+  ret <16 x float> %tmp4;
+}
+
+define <16 x float> @test_min_v16f32(<16 x float>* %a_ptr, <16 x float> %b)  {
+; CHECK_UNSAFE-LABEL: test_min_v16f32:
+; CHECK_UNSAFE:       # BB#0:
+; CHECK_UNSAFE-NEXT:    vminps (%rdi), %zmm0, %zmm0
+; CHECK_UNSAFE-NEXT:    retq
+;
+; CHECK-LABEL: test_min_v16f32:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vmovaps (%rdi), %zmm1
+; CHECK-NEXT:    vminps %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    retq
+  %a = load <16 x float>, <16 x float>* %a_ptr
+  %tmp = fcmp fast olt <16 x float> %a, %b
+  %tmp4 = select <16 x i1> %tmp, <16 x float> %a, <16 x float> %b
+  ret <16 x float> %tmp4;
+}
+
+define <8 x double> @test_max_v8f64(<8 x double> * %a_ptr, <8 x double> %b)  {
+; CHECK_UNSAFE-LABEL: test_max_v8f64:
+; CHECK_UNSAFE:       # BB#0:
+; CHECK_UNSAFE-NEXT:    vmaxpd (%rdi), %zmm0, %zmm0
+; CHECK_UNSAFE-NEXT:    retq
+;
+; CHECK-LABEL: test_max_v8f64:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vmovapd (%rdi), %zmm1
+; CHECK-NEXT:    vmaxpd %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    retq
+  %a = load <8 x double>, <8 x double>* %a_ptr
+  %tmp = fcmp fast ogt <8 x double> %a, %b
+  %tmp4 = select <8 x i1> %tmp, <8 x double> %a, <8 x double> %b
+  ret <8 x double> %tmp4;
+}
+
+define <8 x double> @test_min_v8f64(<8 x double>* %a_ptr, <8 x double> %b)  {
+; CHECK_UNSAFE-LABEL: test_min_v8f64:
+; CHECK_UNSAFE:       # BB#0:
+; CHECK_UNSAFE-NEXT:    vminpd (%rdi), %zmm0, %zmm0
+; CHECK_UNSAFE-NEXT:    retq
+;
+; CHECK-LABEL: test_min_v8f64:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vmovapd (%rdi), %zmm1
+; CHECK-NEXT:    vminpd %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    retq
+  %a = load <8 x double>, <8 x double>* %a_ptr
+  %tmp = fcmp fast olt <8 x double> %a, %b
+  %tmp4 = select <8 x i1> %tmp, <8 x double> %a, <8 x double> %b
+  ret <8 x double> %tmp4;
+}