[X86] Add isel patterns for memory forms of FMA3 intrinsic instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@312309 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrFMA.td b/lib/Target/X86/X86InstrFMA.td
index 9e554db918c64c8f1487582678ab036e695b4de4..fda45907eef7e529bc349c7f27408b999caab07c 100644 (file)
--- a/lib/Target/X86/X86InstrFMA.td
+++ b/lib/Target/X86/X86InstrFMA.td
@@ -242,6 +242,7 @@ multiclass fma3s<bits<8> opc132, bits<8> opc213, bits<8> opc231,
   // This is because src1 is tied to dest, and the scalar intrinsics
   // require the pass-through values to come from the first source
   // operand, not the second.
+  // TODO: Use AVX512 instructions when possible.
   let Predicates = [HasFMA] in {
     def : Pat<(IntF32 VR128:$src1, VR128:$src2, VR128:$src3),
               (!cast<Instruction>(NAME#"213SSr_Int")
@@ -250,6 +251,22 @@ multiclass fma3s<bits<8> opc132, bits<8> opc213, bits<8> opc231,
     def : Pat<(IntF64 VR128:$src1, VR128:$src2, VR128:$src3),
               (!cast<Instruction>(NAME#"213SDr_Int")
                VR128:$src1, VR128:$src2, VR128:$src3)>;
+
+    def : Pat<(IntF32 VR128:$src1, VR128:$src2, sse_load_f32:$src3),
+              (!cast<Instruction>(NAME#"213SSm_Int")
+               VR128:$src1, VR128:$src2, sse_load_f32:$src3)>;
+
+    def : Pat<(IntF64 VR128:$src1, VR128:$src2, sse_load_f64:$src3),
+              (!cast<Instruction>(NAME#"213SDm_Int")
+               VR128:$src1, VR128:$src2, sse_load_f64:$src3)>;
+
+    def : Pat<(IntF32 VR128:$src1, sse_load_f32:$src3, VR128:$src2),
+              (!cast<Instruction>(NAME#"132SSm_Int")
+               VR128:$src1, VR128:$src2, sse_load_f32:$src3)>;
+
+    def : Pat<(IntF64 VR128:$src1, sse_load_f64:$src3, VR128:$src2),
+              (!cast<Instruction>(NAME#"132SDm_Int")
+               VR128:$src1, VR128:$src2, sse_load_f64:$src3)>;
   }
 }
 

diff --git a/test/CodeGen/X86/fma-intrinsics-x86.ll b/test/CodeGen/X86/fma-intrinsics-x86.ll
index fff3a49bea2fe405c25f1458e0f810bfbf326c6d..d30f1c9899e02530608b1388cb32b7c54aaf363d 100644 (file)
--- a/test/CodeGen/X86/fma-intrinsics-x86.ll
+++ b/test/CodeGen/X86/fma-intrinsics-x86.ll
@@ -10,9 +10,9 @@ define <4 x float> @test_x86_fma_vfmadd_ss(<4 x float> %a0, <4 x float> %a1, <4
 ; CHECK-LABEL: test_x86_fma_vfmadd_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmadd132ss     (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmadd213ss     (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0
 ;
@@ -27,9 +27,9 @@ define <4 x float> @test_x86_fma_vfmadd_bac_ss(<4 x float> %a0, <4 x float> %a1,
 ; CHECK-LABEL: test_x86_fma_vfmadd_bac_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmadd132ss (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmadd213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmadd213ss %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovaps   %xmm1, %xmm0
@@ -45,9 +45,9 @@ define <2 x double> @test_x86_fma_vfmadd_sd(<2 x double> %a0, <2 x double> %a1,
 ; CHECK-LABEL: test_x86_fma_vfmadd_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmadd132sd (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmadd213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0
 ;
@@ -62,9 +62,9 @@ define <2 x double> @test_x86_fma_vfmadd_bac_sd(<2 x double> %a0, <2 x double> %
 ; CHECK-LABEL: test_x86_fma_vfmadd_bac_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmadd132sd (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmadd213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmadd213sd %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovapd   %xmm1, %xmm0
@@ -154,9 +154,9 @@ define <4 x float> @test_x86_fma_vfmsub_ss(<4 x float> %a0, <4 x float> %a1, <4
 ; CHECK-LABEL: test_x86_fma_vfmsub_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmsub132ss (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmsub213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmsub213ss %xmm2, %xmm1, %xmm0
 ;
@@ -171,9 +171,9 @@ define <4 x float> @test_x86_fma_vfmsub_bac_ss(<4 x float> %a0, <4 x float> %a1,
 ; CHECK-LABEL: test_x86_fma_vfmsub_bac_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmsub132ss (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmsub213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmsub213ss %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovaps   %xmm1, %xmm0
@@ -190,9 +190,9 @@ define <2 x double> @test_x86_fma_vfmsub_sd(<2 x double> %a0, <2 x double> %a1,
 ; CHECK-LABEL: test_x86_fma_vfmsub_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmsub132sd (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmsub213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmsub213sd %xmm2, %xmm1, %xmm0
 ;
@@ -207,9 +207,9 @@ define <2 x double> @test_x86_fma_vfmsub_bac_sd(<2 x double> %a0, <2 x double> %
 ; CHECK-LABEL: test_x86_fma_vfmsub_bac_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfmsub132sd (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfmsub213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfmsub213sd %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovapd   %xmm1, %xmm0
@@ -299,9 +299,9 @@ define <4 x float> @test_x86_fma_vfnmadd_ss(<4 x float> %a0, <4 x float> %a1, <4
 ; CHECK-LABEL: test_x86_fma_vfnmadd_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmadd132ss (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmadd213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmadd213ss %xmm2, %xmm1, %xmm0
 ;
@@ -316,9 +316,9 @@ define <4 x float> @test_x86_fma_vfnmadd_bac_ss(<4 x float> %a0, <4 x float> %a1
 ; CHECK-LABEL: test_x86_fma_vfnmadd_bac_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmadd132ss (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmadd213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmadd213ss %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovaps   %xmm1, %xmm0
@@ -335,9 +335,9 @@ define <2 x double> @test_x86_fma_vfnmadd_sd(<2 x double> %a0, <2 x double> %a1,
 ; CHECK-LABEL: test_x86_fma_vfnmadd_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmadd132sd (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmadd213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmadd213sd %xmm2, %xmm1, %xmm0
 ;
@@ -352,9 +352,9 @@ define <2 x double> @test_x86_fma_vfnmadd_bac_sd(<2 x double> %a0, <2 x double>
 ; CHECK-LABEL: test_x86_fma_vfnmadd_bac_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmadd132sd (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmadd213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmadd213sd %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovapd   %xmm1, %xmm0
@@ -444,9 +444,9 @@ define <4 x float> @test_x86_fma_vfnmsub_ss(<4 x float> %a0, <4 x float> %a1, <4
 ; CHECK-LABEL: test_x86_fma_vfnmsub_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmsub132ss (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmsub213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmsub213ss %xmm2, %xmm1, %xmm0
 ;
@@ -461,9 +461,9 @@ define <4 x float> @test_x86_fma_vfnmsub_bac_ss(<4 x float> %a0, <4 x float> %a1
 ; CHECK-LABEL: test_x86_fma_vfnmsub_bac_ss:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmsub132ss (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovaps {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmsub213ss (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmsub213ss %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovaps   %xmm1, %xmm0
@@ -480,9 +480,9 @@ define <2 x double> @test_x86_fma_vfnmsub_sd(<2 x double> %a0, <2 x double> %a1,
 ; CHECK-LABEL: test_x86_fma_vfnmsub_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmsub132sd (%rdx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rcx\), %xmm0|\(%rdx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmsub213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmsub213sd %xmm2, %xmm1, %xmm0
 ;
@@ -497,9 +497,9 @@ define <2 x double> @test_x86_fma_vfnmsub_bac_sd(<2 x double> %a0, <2 x double>
 ; CHECK-LABEL: test_x86_fma_vfnmsub_bac_sd:
 ; CHECK-NEXT:  # BB#0:
 ;
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%r8\), %xmm1}}
-; CHECK-FMA-WIN-NEXT: vfnmsub132sd (%rcx), %xmm1, %xmm0
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vmovapd {{\(%rdx\), %xmm0|\(%rcx\), %xmm1}}
+; CHECK-FMA-WIN-NEXT: vfnmsub213sd (%r8), %xmm1, %xmm0
 ;
 ; CHECK-FMA-NEXT:    vfnmsub213sd %xmm2, %xmm0, %xmm1
 ; CHECK-FMA-NEXT:    vmovapd   %xmm1, %xmm0