[x86] enable machine combiner reassociations for scalar single-precision minimums
llvm-svn: 245166
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index c2ff910..bd3a3e5 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -6385,11 +6385,17 @@
// TODO: There are many more machine instruction opcodes to match:
// 1. Other data types (integer, vectors)
// 2. Other math / logic operations (and, or)
+// 3. Other forms of the same operation (intrinsics and other variants)
static bool isAssociativeAndCommutative(const MachineInstr &Inst) {
switch (Inst.getOpcode()) {
case X86::IMUL16rr:
case X86::IMUL32rr:
case X86::IMUL64rr:
+ // Normal min/max instructions are not commutative because of NaN and signed
+ // zero semantics, but these are. Thus, there's no need to check for global
+ // relaxed math; the instructions themselves have the properties we need.
+ case X86::MINCSSrr:
+ case X86::VMINCSSrr:
return true;
case X86::ADDPDrr:
case X86::ADDPSrr:
diff --git a/llvm/test/CodeGen/X86/machine-combiner.ll b/llvm/test/CodeGen/X86/machine-combiner.ll
index 900a43f..3b5e6d2 100644
--- a/llvm/test/CodeGen/X86/machine-combiner.ll
+++ b/llvm/test/CodeGen/X86/machine-combiner.ll
@@ -358,21 +358,21 @@
ret <4 x double> %t2
}
-; TODO: Verify that SSE and AVX scalar single-precision minimum ops are reassociated.
+; Verify that SSE and AVX scalar single-precision minimum ops are reassociated.
define float @reassociate_mins_single(float %x0, float %x1, float %x2, float %x3) {
; SSE-LABEL: reassociate_mins_single:
; SSE: # BB#0:
; SSE-NEXT: divss %xmm1, %xmm0
+; SSE-NEXT: minss %xmm3, %xmm2
; SSE-NEXT: minss %xmm2, %xmm0
-; SSE-NEXT: minss %xmm3, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: reassociate_mins_single:
; AVX: # BB#0:
; AVX-NEXT: vdivss %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vminss %xmm0, %xmm2, %xmm0
-; AVX-NEXT: vminss %xmm0, %xmm3, %xmm0
+; AVX-NEXT: vminss %xmm3, %xmm2, %xmm1
+; AVX-NEXT: vminss %xmm1, %xmm0, %xmm0
; AVX-NEXT: retq
%t0 = fdiv float %x0, %x1
%cmp1 = fcmp olt float %x2, %t0