[DAG] Fix Big Endian in Load-Store forwarding

[niravd]

Summary

Correct offset calculation in load-store forwarding for big-endian
targets.

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::ForwardStoreValueToDirectLoad(LoadSDNode *LD) {
   bool STCoversLD =
       BasePtrST.equalBaseIndex(BasePtrLD, DAG, Offset) && (Offset >= 0) &&
       (Offset * 8 <= LDMemType.getSizeInBits()) &&
       (Offset * 8 + LDMemType.getSizeInBits() <= STMemType.getSizeInBits());
 
   if (!STCoversLD)
     return SDValue();
 
+  // Normalize for Endianness.
+  if (DAG.getDataLayout().isBigEndian())
+    Offset =
+        (STMemType.getSizeInBits() - LDMemType.getSizeInBits()) / 8 - Offset;
+
   // Memory as copy space (potentially masked).
   if (Offset == 0 && LDType == STType && STMemType == LDMemType) {
     // Simple case: Direct non-truncating forwarding
     if (LDType.getSizeInBits() == LDMemType.getSizeInBits())
       return CombineTo(LD, ST->getValue(), Chain);
     // Can we model the truncate and extension with an and mask?
     if (STType.isInteger() && LDMemType.isInteger() && !STType.isVector() &&
         !LDMemType.isVector() && LD->getExtensionType() != ISD::SEXTLOAD) {

llvm/test/CodeGen/Mips/cconv/vector.ll

 ; MIPS32R5EB-NEXT:    copy_s.w $2, $w0[1]
 ; MIPS32R5EB-NEXT:    copy_s.w $3, $w0[3]
 ; MIPS32R5EB-NEXT:    move $sp, $fp
 ; MIPS32R5EB-NEXT:    lw $fp, 44($sp) # 4-byte Folded Reload
 ; MIPS32R5EB-NEXT:    addiu $sp, $sp, 48
 ; MIPS32R5EB-NEXT:    jr $ra
 ; MIPS32R5EB-NEXT:    nop
 ;
-; MIPS64R5-LABEL: i32_2:
-; MIPS64R5:       # %bb.0:
-; MIPS64R5-NEXT:    daddiu $sp, $sp, -32
-; MIPS64R5-NEXT:    .cfi_def_cfa_offset 32
-; MIPS64R5-NEXT:    sd $5, 16($sp)
-; MIPS64R5-NEXT:    sd $4, 24($sp)
-; MIPS64R5-NEXT:    ldi.b $w0, 0
-; MIPS64R5-NEXT:    lw $1, 20($sp)
-; MIPS64R5-NEXT:    move.v $w1, $w0
-; MIPS64R5-NEXT:    insert.d $w1[0], $5
-; MIPS64R5-NEXT:    insert.d $w1[1], $1
-; MIPS64R5-NEXT:    insert.d $w0[0], $4
-; MIPS64R5-NEXT:    lw $1, 28($sp)
-; MIPS64R5-NEXT:    insert.d $w0[1], $1
-; MIPS64R5-NEXT:    addv.d $w0, $w0, $w1
-; MIPS64R5-NEXT:    copy_s.d $1, $w0[0]
-; MIPS64R5-NEXT:    copy_s.d $2, $w0[1]
-; MIPS64R5-NEXT:    sw $2, 12($sp)
-; MIPS64R5-NEXT:    sw $1, 8($sp)
-; MIPS64R5-NEXT:    ld $2, 8($sp)
-; MIPS64R5-NEXT:    daddiu $sp, $sp, 32
-; MIPS64R5-NEXT:    jr $ra
-; MIPS64R5-NEXT:    nop
+; MIPS64R5EB-LABEL: i32_2:
+; MIPS64R5EB:       # %bb.0:
+; MIPS64R5EB-NEXT:    daddiu $sp, $sp, -32
+; MIPS64R5EB-NEXT:    .cfi_def_cfa_offset 32
+; MIPS64R5EB-NEXT:    sd $5, 16($sp)
+; MIPS64R5EB-NEXT:    sd $4, 24($sp)
+; MIPS64R5EB-NEXT:    ldi.b $w0, 0
+; MIPS64R5EB-NEXT:    lw $1, 16($sp)
+; MIPS64R5EB-NEXT:    move.v $w1, $w0
+; MIPS64R5EB-NEXT:    insert.d $w1[0], $1
+; MIPS64R5EB-NEXT:    insert.d $w1[1], $5
+; MIPS64R5EB-NEXT:    lw $1, 24($sp)
+; MIPS64R5EB-NEXT:    insert.d $w0[0], $1
+; MIPS64R5EB-NEXT:    insert.d $w0[1], $4
+; MIPS64R5EB-NEXT:    addv.d $w0, $w0, $w1
+; MIPS64R5EB-NEXT:    copy_s.d $1, $w0[0]
+; MIPS64R5EB-NEXT:    copy_s.d $2, $w0[1]
+; MIPS64R5EB-NEXT:    sw $2, 12($sp)
+; MIPS64R5EB-NEXT:    sw $1, 8($sp)
+; MIPS64R5EB-NEXT:    ld $2, 8($sp)
+; MIPS64R5EB-NEXT:    daddiu $sp, $sp, 32
+; MIPS64R5EB-NEXT:    jr $ra
+; MIPS64R5EB-NEXT:    nop
 ;
 ; MIPS32R5EL-LABEL: i32_2:
 ; MIPS32R5EL:       # %bb.0:
 ; MIPS32R5EL-NEXT:    addiu $sp, $sp, -48
 ; MIPS32R5EL-NEXT:    .cfi_def_cfa_offset 48
 ; MIPS32R5EL-NEXT:    sw $fp, 44($sp) # 4-byte Folded Spill
 ; MIPS32R5EL-NEXT:    .cfi_offset 30, -4
 ; MIPS32R5EL-NEXT:    move $fp, $sp

 ; MIPS32R5EL-NEXT:    addv.d $w0, $w1, $w0
 ; MIPS32R5EL-NEXT:    copy_s.w $2, $w0[0]
 ; MIPS32R5EL-NEXT:    copy_s.w $3, $w0[2]
 ; MIPS32R5EL-NEXT:    move $sp, $fp
 ; MIPS32R5EL-NEXT:    lw $fp, 44($sp) # 4-byte Folded Reload
 ; MIPS32R5EL-NEXT:    addiu $sp, $sp, 48
 ; MIPS32R5EL-NEXT:    jr $ra
 ; MIPS32R5EL-NEXT:    nop
+;
+; MIPS64R5EL-LABEL: i32_2:
+; MIPS64R5EL:       # %bb.0:
+; MIPS64R5EL-NEXT:    daddiu $sp, $sp, -32
+; MIPS64R5EL-NEXT:    .cfi_def_cfa_offset 32
+; MIPS64R5EL-NEXT:    sd $5, 16($sp)
+; MIPS64R5EL-NEXT:    sd $4, 24($sp)
+; MIPS64R5EL-NEXT:    ldi.b $w0, 0
+; MIPS64R5EL-NEXT:    lw $1, 20($sp)
+; MIPS64R5EL-NEXT:    move.v $w1, $w0
+; MIPS64R5EL-NEXT:    insert.d $w1[0], $5
+; MIPS64R5EL-NEXT:    insert.d $w1[1], $1
+; MIPS64R5EL-NEXT:    insert.d $w0[0], $4
+; MIPS64R5EL-NEXT:    lw $1, 28($sp)
+; MIPS64R5EL-NEXT:    insert.d $w0[1], $1
+; MIPS64R5EL-NEXT:    addv.d $w0, $w0, $w1
+; MIPS64R5EL-NEXT:    copy_s.d $1, $w0[0]
+; MIPS64R5EL-NEXT:    copy_s.d $2, $w0[1]
+; MIPS64R5EL-NEXT:    sw $2, 12($sp)
+; MIPS64R5EL-NEXT:    sw $1, 8($sp)
+; MIPS64R5EL-NEXT:    ld $2, 8($sp)
+; MIPS64R5EL-NEXT:    daddiu $sp, $sp, 32
+; MIPS64R5EL-NEXT:    jr $ra
+; MIPS64R5EL-NEXT:    nop
   %1 = add <2 x i32> %a, %b
   ret <2 x i32> %1
 }
 
 define <4 x i32> @i32_4(<4 x i32> %a, <4 x i32> %b) {
 ; MIPS32-LABEL: i32_4:
 ; MIPS32:       # %bb.0:
 ; MIPS32-NEXT:    lw $1, 20($sp)

llvm/test/CodeGen/PowerPC/big-endian-store-forward.ll

+; RUN: llc -mtriple=powerpc64-unknown-linux-gnu < %s | FileCheck %s
+
+; The load is to the high byte of the 2-byte store
+@g = global i8 -75
+
+define void @f(i16 %v) {
+; CHECK-LABEL: f
+; CHECK: sth 3, -2(1)
+; CHECK: lbz 3, -2(1)
+  %p32 = alloca i16
+  store i16 %v, i16* %p32
+  %p16 = bitcast i16* %p32 to i8*
+  %tmp = load i8, i8* %p16
+  store i8 %tmp, i8* @g
+  ret void
+}