[LoongArch][ISel] Check the number of sign bits in `PatGprGpr_32` (#107432)

After https://github.com/llvm/llvm-project/pull/92205, LoongArch ISel
selects `div.w` for `trunc i64 (sdiv i64 3202030857, (sext i32 X to
i64)) to i32`. It is incorrect since `3202030857` is not a signed 32-bit
constant. It will produce wrong result when `X == 2`:
https://alive2.llvm.org/ce/z/pzfGZZ

This patch adds additional `sexti32` checks to operands of
`PatGprGpr_32`.
Alive2 proof: https://alive2.llvm.org/ce/z/AkH5Mp

Fix #107414.

(cherry picked from commit a111f9119a5ec77c19a514ec09454218f739454f)

diff --git a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
index ef647a42778737c1d8fbe9dde3359428ae3f7914..339d50bd8192176ea489be3708b931296f25cc98 100644 (file)
--- a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
@@ -1065,10 +1065,13 @@ def RDTIME_D : RDTIME_2R<0x00006800>;
 
 /// Generic pattern classes
 
+def assertsexti32 : PatFrag<(ops node:$src), (assertsext node:$src), [{
+  return cast<VTSDNode>(N->getOperand(1))->getVT().bitsLE(MVT::i32);
+}]>;
 class PatGprGpr<SDPatternOperator OpNode, LAInst Inst>
     : Pat<(OpNode GPR:$rj, GPR:$rk), (Inst GPR:$rj, GPR:$rk)>;
 class PatGprGpr_32<SDPatternOperator OpNode, LAInst Inst>
-    : Pat<(sext_inreg (OpNode GPR:$rj, GPR:$rk), i32), (Inst GPR:$rj, GPR:$rk)>;
+    : Pat<(sext_inreg (OpNode (assertsexti32 GPR:$rj), (assertsexti32 GPR:$rk)), i32), (Inst GPR:$rj, GPR:$rk)>;
 class PatGpr<SDPatternOperator OpNode, LAInst Inst>
     : Pat<(OpNode GPR:$rj), (Inst GPR:$rj)>;
 

diff --git a/llvm/test/CodeGen/LoongArch/ir-instruction/sdiv-udiv-srem-urem.ll b/llvm/test/CodeGen/LoongArch/ir-instruction/sdiv-udiv-srem-urem.ll
index ab3eec240db3c14278183a1520b4fd7c320855bf..c22acdb49690715f9e29e489c5e18e13b3653e93 100644 (file)
--- a/llvm/test/CodeGen/LoongArch/ir-instruction/sdiv-udiv-srem-urem.ll
+++ b/llvm/test/CodeGen/LoongArch/ir-instruction/sdiv-udiv-srem-urem.ll
@@ -191,7 +191,8 @@ define signext i32 @sdiv_si32_ui32_ui32(i32 %a, i32 %b) {
 ; LA64:       # %bb.0: # %entry
 ; LA64-NEXT:    addi.w $a1, $a1, 0
 ; LA64-NEXT:    addi.w $a0, $a0, 0
-; LA64-NEXT:    div.w $a0, $a0, $a1
+; LA64-NEXT:    div.d $a0, $a0, $a1
+; LA64-NEXT:    addi.w $a0, $a0, 0
 ; LA64-NEXT:    ret
 ;
 ; LA32-TRAP-LABEL: sdiv_si32_ui32_ui32:
@@ -207,11 +208,12 @@ define signext i32 @sdiv_si32_ui32_ui32(i32 %a, i32 %b) {
 ; LA64-TRAP:       # %bb.0: # %entry
 ; LA64-TRAP-NEXT:    addi.w $a1, $a1, 0
 ; LA64-TRAP-NEXT:    addi.w $a0, $a0, 0
-; LA64-TRAP-NEXT:    div.w $a0, $a0, $a1
+; LA64-TRAP-NEXT:    div.d $a0, $a0, $a1
 ; LA64-TRAP-NEXT:    bnez $a1, .LBB5_2
 ; LA64-TRAP-NEXT:  # %bb.1: # %entry
 ; LA64-TRAP-NEXT:    break 7
 ; LA64-TRAP-NEXT:  .LBB5_2: # %entry
+; LA64-TRAP-NEXT:    addi.w $a0, $a0, 0
 ; LA64-TRAP-NEXT:    ret
 entry:
   %r = sdiv i32 %a, %b
@@ -1151,3 +1153,64 @@ entry:
   %r = urem i64 %a, %b
   ret i64 %r
 }
+
+define signext i32 @pr107414(i32 signext %x) {
+; LA32-LABEL: pr107414:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    addi.w $sp, $sp, -16
+; LA32-NEXT:    .cfi_def_cfa_offset 16
+; LA32-NEXT:    st.w $ra, $sp, 12 # 4-byte Folded Spill
+; LA32-NEXT:    .cfi_offset 1, -4
+; LA32-NEXT:    move $a2, $a0
+; LA32-NEXT:    srai.w $a3, $a0, 31
+; LA32-NEXT:    lu12i.w $a0, -266831
+; LA32-NEXT:    ori $a0, $a0, 3337
+; LA32-NEXT:    move $a1, $zero
+; LA32-NEXT:    bl %plt(__divdi3)
+; LA32-NEXT:    ld.w $ra, $sp, 12 # 4-byte Folded Reload
+; LA32-NEXT:    addi.w $sp, $sp, 16
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: pr107414:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    lu12i.w $a1, -266831
+; LA64-NEXT:    ori $a1, $a1, 3337
+; LA64-NEXT:    lu32i.d $a1, 0
+; LA64-NEXT:    div.d $a0, $a1, $a0
+; LA64-NEXT:    addi.w $a0, $a0, 0
+; LA64-NEXT:    ret
+;
+; LA32-TRAP-LABEL: pr107414:
+; LA32-TRAP:       # %bb.0: # %entry
+; LA32-TRAP-NEXT:    addi.w $sp, $sp, -16
+; LA32-TRAP-NEXT:    .cfi_def_cfa_offset 16
+; LA32-TRAP-NEXT:    st.w $ra, $sp, 12 # 4-byte Folded Spill
+; LA32-TRAP-NEXT:    .cfi_offset 1, -4
+; LA32-TRAP-NEXT:    move $a2, $a0
+; LA32-TRAP-NEXT:    srai.w $a3, $a0, 31
+; LA32-TRAP-NEXT:    lu12i.w $a0, -266831
+; LA32-TRAP-NEXT:    ori $a0, $a0, 3337
+; LA32-TRAP-NEXT:    move $a1, $zero
+; LA32-TRAP-NEXT:    bl %plt(__divdi3)
+; LA32-TRAP-NEXT:    ld.w $ra, $sp, 12 # 4-byte Folded Reload
+; LA32-TRAP-NEXT:    addi.w $sp, $sp, 16
+; LA32-TRAP-NEXT:    ret
+;
+; LA64-TRAP-LABEL: pr107414:
+; LA64-TRAP:       # %bb.0: # %entry
+; LA64-TRAP-NEXT:    lu12i.w $a1, -266831
+; LA64-TRAP-NEXT:    ori $a1, $a1, 3337
+; LA64-TRAP-NEXT:    lu32i.d $a1, 0
+; LA64-TRAP-NEXT:    div.d $a1, $a1, $a0
+; LA64-TRAP-NEXT:    bnez $a0, .LBB32_2
+; LA64-TRAP-NEXT:  # %bb.1: # %entry
+; LA64-TRAP-NEXT:    break 7
+; LA64-TRAP-NEXT:  .LBB32_2: # %entry
+; LA64-TRAP-NEXT:    addi.w $a0, $a1, 0
+; LA64-TRAP-NEXT:    ret
+entry:
+  %conv = sext i32 %x to i64
+  %div = sdiv i64 3202030857, %conv
+  %conv1 = trunc i64 %div to i32
+  ret i32 %conv1
+}