From f84fc5ad12591b67af54c5483ff29aba61a7190a Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Thu, 18 Jul 2019 12:48:01 +0000
Subject: [PATCH] [x86] try harder to form LEA from ADD to avoid flag conflicts
 (PR40483)

LEA doesn't affect flags, so use it more liberally to replace an ADD when
we know that the ADD operands affect flags.

In the motivating example from PR40483:
https://bugs.llvm.org/show_bug.cgi?id=40483
...this lets us avoid duplicating a math op just to avoid flag conflict.

As mentioned in the TODO comments, this heuristic can be extended to
fire more often if that leads to more improvements.

Differential Revision: https://reviews.llvm.org/D64707

llvm-svn: 366431
---
 lib/Target/X86/X86ISelDAGToDAG.cpp | 31 ++++++++++++++++++++++++++++++
 test/CodeGen/X86/combine-sbb.ll    | 30 ++++++++++-------------------
 2 files changed, 41 insertions(+), 20 deletions(-)

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index 95d31e62caf..fcf9b4148b4 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -2464,6 +2464,37 @@ bool X86DAGToDAGISel::selectLEAAddr(SDValue N,
       Complexity += 2;
   }
 
+  // Heuristic: try harder to form an LEA from ADD if the operands set flags.
+  // Unlike ADD, LEA does not affect flags, so we will be less likely to require
+  // duplicating flag-producing instructions later in the pipeline.
+  if (N.getOpcode() == ISD::ADD) {
+    auto isMathWithFlags = [](SDValue V) {
+      switch (V.getOpcode()) {
+      case X86ISD::ADD:
+      case X86ISD::SUB:
+      case X86ISD::ADC:
+      case X86ISD::SBB:
+      /* TODO: These opcodes can be added safely, but we may want to justify
+               their inclusion for different reasons (better for reg-alloc).
+      case X86ISD::SMUL:
+      case X86ISD::UMUL:
+      case X86ISD::OR:
+      case X86ISD::XOR:
+      case X86ISD::AND:
+      */
+        // Value 1 is the flag output of the node - verify it's not dead.
+        return !SDValue(V.getNode(), 1).use_empty();
+      default:
+        return false;
+      }
+    };
+    // TODO: This could be an 'or' rather than 'and' to make the transform more
+    //       likely to happen. We might want to factor in whether there's a
+    //       load folding opportunity for the math op that disappears with LEA.
+    if (isMathWithFlags(N.getOperand(0)) && isMathWithFlags(N.getOperand(1)))
+      Complexity++;
+  }
+
   if (AM.Disp)
     Complexity++;
 
diff --git a/test/CodeGen/X86/combine-sbb.ll b/test/CodeGen/X86/combine-sbb.ll
index bba72c56609..6eb0e1e0f0c 100644
--- a/test/CodeGen/X86/combine-sbb.ll
+++ b/test/CodeGen/X86/combine-sbb.ll
@@ -309,35 +309,25 @@ define i32 @PR40483_sub5(i32*, i32) nounwind {
 define i32 @PR40483_sub6(i32*, i32) nounwind {
 ; X86-LABEL: PR40483_sub6:
 ; X86:       # %bb.0:
-; X86-NEXT:    pushl %edi
-; X86-NEXT:    pushl %esi
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
-; X86-NEXT:    movl (%edx), %esi
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
-; X86-NEXT:    movl %esi, %ecx
-; X86-NEXT:    subl %edi, %ecx
+; X86-NEXT:    movl (%edx), %ecx
 ; X86-NEXT:    xorl %eax, %eax
-; X86-NEXT:    subl %edi, %esi
-; X86-NEXT:    movl %esi, (%edx)
+; X86-NEXT:    subl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl %ecx, (%edx)
 ; X86-NEXT:    jae .LBB8_2
 ; X86-NEXT:  # %bb.1:
-; X86-NEXT:    addl %ecx, %ecx
-; X86-NEXT:    movl %ecx, %eax
+; X86-NEXT:    leal (%ecx,%ecx), %eax
 ; X86-NEXT:  .LBB8_2:
-; X86-NEXT:    popl %esi
-; X86-NEXT:    popl %edi
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: PR40483_sub6:
 ; X64:       # %bb.0:
-; X64-NEXT:    movl (%rdi), %ecx
-; X64-NEXT:    movl %ecx, %edx
-; X64-NEXT:    subl %esi, %edx
-; X64-NEXT:    addl %edx, %edx
-; X64-NEXT:    xorl %eax, %eax
-; X64-NEXT:    subl %esi, %ecx
-; X64-NEXT:    movl %ecx, (%rdi)
-; X64-NEXT:    cmovbl %edx, %eax
+; X64-NEXT:    movl (%rdi), %eax
+; X64-NEXT:    xorl %ecx, %ecx
+; X64-NEXT:    subl %esi, %eax
+; X64-NEXT:    movl %eax, (%rdi)
+; X64-NEXT:    leal (%rax,%rax), %eax
+; X64-NEXT:    cmovael %ecx, %eax
 ; X64-NEXT:    retq
   %3 = load i32, i32* %0, align 8
   %4 = tail call { i8, i32 } @llvm.x86.subborrow.32(i8 0, i32 %3, i32 %1)