[ARM] Select VMAXNM and VMINNM regardless of operand order

[olista01]

Summary

Currently, the ARM backend will select the VMAXNM and VMINNM for these C
expressions:

(a < b) ? a : b
(a > b) ? a : b

but not these expressions:

(a > b) ? b : a
(a < b) ? b : a

This patch allows all of these expressions to be matched.

[rengolin]

LGTM, Thanks!

[olista01]

Thanks, committed as r220671.

lib/Target/ARM/ARMISelLowering.cpp

SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
   if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
                                     TrueVal.getValueType() == MVT::f64)) {
     // We can select VMAXNM/VMINNM from a compare followed by a select with the
     // same operands, as follows:
     //   c = fcmp [ogt, olt, ugt, ult] a, b
     //   select c, a, b
     // We only do this in unsafe-fp-math, because signed zeros and NaNs are
     // handled differently than the original code sequence.
-    if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
-        RHS == FalseVal) {
+    if (getTargetMachine().Options.UnsafeFPMath) {
+      if (LHS == TrueVal && RHS == FalseVal) {
         if (CC == ISD::SETOGT || CC == ISD::SETUGT)
           return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
         if (CC == ISD::SETOLT || CC == ISD::SETULT)
           return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
+      } else if (LHS == FalseVal && RHS == TrueVal) {
+        if (CC == ISD::SETOLT || CC == ISD::SETULT)
+          return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
+        if (CC == ISD::SETOGT || CC == ISD::SETUGT)
+          return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
+      }
     }
 
     bool swpCmpOps = false;
     bool swpVselOps = false;
     checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
 
     if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
         CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {

test/CodeGen/ARM/vminmaxnm.ll

 ; RUN: llc < %s -mtriple armv8 -mattr=+neon | FileCheck %s
 ; RUN: llc < %s -mtriple armv8 -mattr=+neon,+fp-armv8 -enable-unsafe-fp-math | FileCheck %s --check-prefix=CHECK-FAST
 
 define <4 x float> @vmaxnmq(<4 x float>* %A, <4 x float>* %B) nounwind {
-; CHECK: vmaxnmq
+; CHECK-LABEL: vmaxnmq:
 ; CHECK: vmaxnm.f32 q{{[0-9]+}}, q{{[0-9]+}}, q{{[0-9]+}}
   %tmp1 = load <4 x float>* %A
   %tmp2 = load <4 x float>* %B
   %tmp3 = call <4 x float> @llvm.arm.neon.vmaxnm.v4f32(<4 x float> %tmp1, <4 x float> %tmp2)
   ret <4 x float> %tmp3
 }
 
 define <2 x float> @vmaxnmd(<2 x float>* %A, <2 x float>* %B) nounwind {
-; CHECK: vmaxnmd
+; CHECK-LABEL: vmaxnmd:
 ; CHECK: vmaxnm.f32 d{{[0-9]+}}, d{{[0-9]+}}, d{{[0-9]+}}
   %tmp1 = load <2 x float>* %A
   %tmp2 = load <2 x float>* %B
   %tmp3 = call <2 x float> @llvm.arm.neon.vmaxnm.v2f32(<2 x float> %tmp1, <2 x float> %tmp2)
   ret <2 x float> %tmp3
 }
 
 define <4 x float> @vminnmq(<4 x float>* %A, <4 x float>* %B) nounwind {
-; CHECK: vminnmq
+; CHECK-LABEL: vminnmq:
 ; CHECK: vminnm.f32 q{{[0-9]+}}, q{{[0-9]+}}, q{{[0-9]+}}
   %tmp1 = load <4 x float>* %A
   %tmp2 = load <4 x float>* %B
   %tmp3 = call <4 x float> @llvm.arm.neon.vminnm.v4f32(<4 x float> %tmp1, <4 x float> %tmp2)
   ret <4 x float> %tmp3
 }
 
 define <2 x float> @vminnmd(<2 x float>* %A, <2 x float>* %B) nounwind {
-; CHECK: vminnmd
+; CHECK-LABEL: vminnmd:
 ; CHECK: vminnm.f32 d{{[0-9]+}}, d{{[0-9]+}}, d{{[0-9]+}}
   %tmp1 = load <2 x float>* %A
   %tmp2 = load <2 x float>* %B
   %tmp3 = call <2 x float> @llvm.arm.neon.vminnm.v2f32(<2 x float> %tmp1, <2 x float> %tmp2)
   ret <2 x float> %tmp3
 }
 
 define float @fp-armv8_vminnm_o(float %a, float %b) {
-; CHECK-FAST: fp-armv8_vminnm_o
+; CHECK-FAST-LABEL: "fp-armv8_vminnm_o":
 ; CHECK-FAST-NOT: vcmp
 ; CHECK-FAST: vminnm.f32
-; CHECK: fp-armv8_vminnm_o
+; CHECK-LABEL: "fp-armv8_vminnm_o":
 ; CHECK-NOT: vminnm.f32
   %cmp = fcmp olt float %a, %b
   %cond = select i1 %cmp, float %a, float %b
   ret float %cond
 }
 
+define float @fp-armv8_vminnm_o_rev(float %a, float %b) {
+; CHECK-FAST-LABEL: "fp-armv8_vminnm_o_rev":
+; CHECK-FAST-NOT: vcmp
+; CHECK-FAST: vminnm.f32
+; CHECK-LABEL: "fp-armv8_vminnm_o_rev":
+; CHECK-NOT: vminnm.f32
+  %cmp = fcmp ogt float %a, %b
+  %cond = select i1 %cmp, float %b, float %a
+  ret float %cond
+}
+
 define float @fp-armv8_vminnm_u(float %a, float %b) {
-; CHECK-FAST: fp-armv8_vminnm_u
+; CHECK-FAST-LABEL: "fp-armv8_vminnm_u":
 ; CHECK-FAST-NOT: vcmp
 ; CHECK-FAST: vminnm.f32
-; CHECK: fp-armv8_vminnm_u
+; CHECK-LABEL: "fp-armv8_vminnm_u":
 ; CHECK-NOT: vminnm.f32
   %cmp = fcmp ult float %a, %b
   %cond = select i1 %cmp, float %a, float %b
   ret float %cond
 }
 
+define float @fp-armv8_vminnm_u_rev(float %a, float %b) {
+; CHECK-FAST-LABEL: "fp-armv8_vminnm_u_rev":
+; CHECK-FAST-NOT: vcmp
+; CHECK-FAST: vminnm.f32
+; CHECK-LABEL: "fp-armv8_vminnm_u_rev":
+; CHECK-NOT: vminnm.f32
+  %cmp = fcmp ugt float %a, %b
+  %cond = select i1 %cmp, float %b, float %a
+  ret float %cond
+}
+
 define float @fp-armv8_vmaxnm_o(float %a, float %b) {
-; CHECK-FAST: fp-armv8_vmaxnm_o
+; CHECK-FAST-LABEL: "fp-armv8_vmaxnm_o":
 ; CHECK-FAST-NOT: vcmp
 ; CHECK-FAST: vmaxnm.f32
-; CHECK: fp-armv8_vmaxnm_o
+; CHECK-LABEL: "fp-armv8_vmaxnm_o":
 ; CHECK-NOT: vmaxnm.f32
   %cmp = fcmp ogt float %a, %b
   %cond = select i1 %cmp, float %a, float %b
   ret float %cond
 }
 
+define float @fp-armv8_vmaxnm_o_rev(float %a, float %b) {
+; CHECK-FAST-LABEL: "fp-armv8_vmaxnm_o_rev":
+; CHECK-FAST-NOT: vcmp
+; CHECK-FAST: vmaxnm.f32
+; CHECK-LABEL: "fp-armv8_vmaxnm_o_rev":
+; CHECK-NOT: vmaxnm.f32
+  %cmp = fcmp olt float %a, %b
+  %cond = select i1 %cmp, float %b, float %a
+  ret float %cond
+}
+
 define float @fp-armv8_vmaxnm_u(float %a, float %b) {
-; CHECK-FAST: fp-armv8_vmaxnm_u
+; CHECK-FAST-LABEL: "fp-armv8_vmaxnm_u":
 ; CHECK-FAST-NOT: vcmp
 ; CHECK-FAST: vmaxnm.f32
-; CHECK: fp-armv8_vmaxnm_u
+; CHECK-LABEL: "fp-armv8_vmaxnm_u":
 ; CHECK-NOT: vmaxnm.f32
   %cmp = fcmp ugt float %a, %b
   %cond = select i1 %cmp, float %a, float %b
   ret float %cond
 }
 
+define float @fp-armv8_vmaxnm_u_rev(float %a, float %b) {
+; CHECK-FAST-LABEL: "fp-armv8_vmaxnm_u_rev":
+; CHECK-FAST-NOT: vcmp
+; CHECK-FAST: vmaxnm.f32
+; CHECK-LABEL: "fp-armv8_vmaxnm_u_rev":
+; CHECK-NOT: vmaxnm.f32
+  %cmp = fcmp ult float %a, %b
+  %cond = select i1 %cmp, float %b, float %a
+  ret float %cond
+}
+
 
 declare <4 x float> @llvm.arm.neon.vminnm.v4f32(<4 x float>, <4 x float>) nounwind readnone
 declare <2 x float> @llvm.arm.neon.vminnm.v2f32(<2 x float>, <2 x float>) nounwind readnone
 declare <4 x float> @llvm.arm.neon.vmaxnm.v4f32(<4 x float>, <4 x float>) nounwind readnone
 declare <2 x float> @llvm.arm.neon.vmaxnm.v2f32(<2 x float>, <2 x float>) nounwind readnone