[X86] Add a missing FMA3 scalar intrinsic pattern.

This allows us to use 231 form to fold an insertelement on the add input to the fma. There is technically no software intrinsic that can use this until AVX512F, but it can be manually built up from other intrinsics.

llvm-svn: 337223

diff --git a/llvm/lib/Target/X86/X86InstrFMA.td b/llvm/lib/Target/X86/X86InstrFMA.td
index 7a35d07495ce02d168cab63f380f0bade39fd351..a559f62c8f38701a3f96141d3c33be8aa16e0b1e 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrFMA.td
+++ b/llvm/lib/Target/X86/X86InstrFMA.td
@@ -340,6 +340,13 @@ multiclass scalar_fma_patterns<SDNode Op, string Prefix, string Suffix,
                VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
                (VT (COPY_TO_REGCLASS RC:$src3, VR128)))>;
 
+    def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
+                (Op RC:$src2, RC:$src3,
+                    (EltVT (extractelt (VT VR128:$src1), (iPTR 0)))))))),
+              (!cast<Instruction>(Prefix#"231"#Suffix#"r_Int")
+               VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
+               (VT (COPY_TO_REGCLASS RC:$src3, VR128)))>;
+
     def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
                 (Op RC:$src2,
                     (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),

diff --git a/llvm/test/CodeGen/X86/fma-intrinsics-x86.ll b/llvm/test/CodeGen/X86/fma-intrinsics-x86.ll
index ec4936cf4cea142f3c3c627324ddb4d4976398c4..055e2bb16c142244446eeaae7da39fd449ca8a6c 100644 (file)
--- a/llvm/test/CodeGen/X86/fma-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/fma-intrinsics-x86.ll
@@ -64,6 +64,37 @@ define <4 x float> @test_x86_fma_vfmadd_bac_ss(<4 x float> %a0, <4 x float> %a1,
   ret <4 x float> %5
 }
 
+define <4 x float> @test_x86_fma_vfmadd_ss_231(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 {
+; CHECK-FMA-LABEL: test_x86_fma_vfmadd_ss_231:
+; CHECK-FMA:       # %bb.0:
+; CHECK-FMA-NEXT:    vfmadd231ss %xmm1, %xmm0, %xmm2 # encoding: [0xc4,0xe2,0x79,0xb9,0xd1]
+; CHECK-FMA-NEXT:    # xmm2 = (xmm0 * xmm1) + xmm2
+; CHECK-FMA-NEXT:    vmovaps %xmm2, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc2]
+; CHECK-FMA-NEXT:    retq # encoding: [0xc3]
+;
+; CHECK-AVX512VL-LABEL: test_x86_fma_vfmadd_ss_231:
+; CHECK-AVX512VL:       # %bb.0:
+; CHECK-AVX512VL-NEXT:    vfmadd231ss %xmm1, %xmm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xb9,0xd1]
+; CHECK-AVX512VL-NEXT:    # xmm2 = (xmm0 * xmm1) + xmm2
+; CHECK-AVX512VL-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
+; CHECK-AVX512VL-NEXT:    retq # encoding: [0xc3]
+;
+; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_ss_231:
+; CHECK-FMA-WIN:       # %bb.0:
+; CHECK-FMA-WIN-NEXT:    vmovaps (%r8), %xmm0 # encoding: [0xc4,0xc1,0x78,0x28,0x00]
+; CHECK-FMA-WIN-NEXT:    vmovss (%rcx), %xmm1 # encoding: [0xc5,0xfa,0x10,0x09]
+; CHECK-FMA-WIN-NEXT:    # xmm1 = mem[0],zero,zero,zero
+; CHECK-FMA-WIN-NEXT:    vfmadd231ss (%rdx), %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xb9,0x02]
+; CHECK-FMA-WIN-NEXT:    # xmm0 = (xmm1 * mem) + xmm0
+; CHECK-FMA-WIN-NEXT:    retq # encoding: [0xc3]
+  %1 = extractelement <4 x float> %a0, i64 0
+  %2 = extractelement <4 x float> %a1, i64 0
+  %3 = extractelement <4 x float> %a2, i64 0
+  %4 = call float @llvm.fma.f32(float %1, float %2, float %3)
+  %5 = insertelement <4 x float> %a2, float %4, i64 0
+  ret <4 x float> %5
+}
+
 define <2 x double> @test_x86_fma_vfmadd_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 {
 ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_sd:
 ; CHECK-FMA:       # %bb.0: