[x86] enable machine combiner reassociations for 128-bit vector min/max

[oota-llvm.git] / test / CodeGen / X86 / machine-combiner.ll

diff --git a/test/CodeGen/X86/machine-combiner.ll b/test/CodeGen/X86/machine-combiner.ll
index e9300b4c6e9c522a041581619c13c492b2e6da62..371747479e58a4ba77c549df1173e018f6783246 100644 (file)
--- a/test/CodeGen/X86/machine-combiner.ll
+++ b/test/CodeGen/X86/machine-combiner.ll
@@ -454,3 +454,99 @@ define double @reassociate_maxs_double(double %x0, double %x1, double %x2, doubl
   ret double %sel2
 }
 
+; Verify that SSE and AVX 128-bit vector single-precision minimum ops are reassociated.
+
+define <4 x float> @reassociate_mins_v4f32(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) {
+; SSE-LABEL: reassociate_mins_v4f32:
+; SSE:       # BB#0:
+; SSE-NEXT:    addps %xmm1, %xmm0
+; SSE-NEXT:    minps %xmm3, %xmm2
+; SSE-NEXT:    minps %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_mins_v4f32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vminps %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vminps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %t0 = fadd <4 x float> %x0, %x1
+  %cmp1 = fcmp olt <4 x float> %x2, %t0
+  %sel1 = select <4 x i1> %cmp1, <4 x float> %x2, <4 x float> %t0
+  %cmp2 = fcmp olt <4 x float> %x3, %sel1
+  %sel2 = select <4 x i1> %cmp2, <4 x float> %x3, <4 x float> %sel1
+  ret <4 x float> %sel2
+}
+
+; Verify that SSE and AVX 128-bit vector single-precision maximum ops are reassociated.
+
+define <4 x float> @reassociate_maxs_v4f32(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) {
+; SSE-LABEL: reassociate_maxs_v4f32:
+; SSE:       # BB#0:
+; SSE-NEXT:    addps %xmm1, %xmm0
+; SSE-NEXT:    maxps %xmm3, %xmm2
+; SSE-NEXT:    maxps %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_maxs_v4f32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vmaxps %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vmaxps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %t0 = fadd <4 x float> %x0, %x1
+  %cmp1 = fcmp ogt <4 x float> %x2, %t0
+  %sel1 = select <4 x i1> %cmp1, <4 x float> %x2, <4 x float> %t0
+  %cmp2 = fcmp ogt <4 x float> %x3, %sel1
+  %sel2 = select <4 x i1> %cmp2, <4 x float> %x3, <4 x float> %sel1
+  ret <4 x float> %sel2
+}
+
+; Verify that SSE and AVX 128-bit vector double-precision minimum ops are reassociated.
+
+define <2 x double> @reassociate_mins_v2f64(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, <2 x double> %x3) {
+; SSE-LABEL: reassociate_mins_v2f64:
+; SSE:       # BB#0:
+; SSE-NEXT:    addpd %xmm1, %xmm0
+; SSE-NEXT:    minpd %xmm3, %xmm2
+; SSE-NEXT:    minpd %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_mins_v2f64:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vminpd %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vminpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %t0 = fadd <2 x double> %x0, %x1
+  %cmp1 = fcmp olt <2 x double> %x2, %t0
+  %sel1 = select <2 x i1> %cmp1, <2 x double> %x2, <2 x double> %t0
+  %cmp2 = fcmp olt <2 x double> %x3, %sel1
+  %sel2 = select <2 x i1> %cmp2, <2 x double> %x3, <2 x double> %sel1
+  ret <2 x double> %sel2
+}
+
+; Verify that SSE and AVX 128-bit vector double-precision maximum ops are reassociated.
+
+define <2 x double> @reassociate_maxs_v2f64(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, <2 x double> %x3) {
+; SSE-LABEL: reassociate_maxs_v2f64:
+; SSE:       # BB#0:
+; SSE-NEXT:    addpd %xmm1, %xmm0
+; SSE-NEXT:    maxpd %xmm3, %xmm2
+; SSE-NEXT:    maxpd %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_maxs_v2f64:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vmaxpd %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %t0 = fadd <2 x double> %x0, %x1
+  %cmp1 = fcmp ogt <2 x double> %x2, %t0
+  %sel1 = select <2 x i1> %cmp1, <2 x double> %x2, <2 x double> %t0
+  %cmp2 = fcmp ogt <2 x double> %x3, %sel1
+  %sel2 = select <2 x i1> %cmp2, <2 x double> %x3, <2 x double> %sel1
+  ret <2 x double> %sel2
+}
+