[llvm] r340303 - [BypassSlowDivision] Teach bypass slow division not to interfere with div by constant where constants have been constant hoisted, but not moved from their basic block
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 21 10:15:34 PDT 2018
Author: ctopper
Date: Tue Aug 21 10:15:33 2018
New Revision: 340303
URL: http://llvm.org/viewvc/llvm-project?rev=340303&view=rev
Log:
[BypassSlowDivision] Teach bypass slow division not to interfere with div by constant where constants have been constant hoisted, but not moved from their basic block
DAGCombiner doesn't pay attention to whether constants are opaque before doing the div by constant optimization. So BypassSlowDivision shouldn't introduce control flow that would make DAGCombiner unable to see an opaque constant. This can occur when a div and rem of the same constant are used in the same basic block. it will be hoisted, but not leave the block.
Longer term we probably need to look into the X86 immediate cost model used by constant hoisting and maybe not mark div/rem immediates for hoisting at all.
This fixes the case from PR38649.
Differential Revision: https://reviews.llvm.org/D51000
Modified:
llvm/trunk/lib/Transforms/Utils/BypassSlowDivision.cpp
llvm/trunk/test/CodeGen/X86/divide-by-constant.ll
Modified: llvm/trunk/lib/Transforms/Utils/BypassSlowDivision.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Utils/BypassSlowDivision.cpp?rev=340303&r1=340302&r2=340303&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Utils/BypassSlowDivision.cpp (original)
+++ llvm/trunk/lib/Transforms/Utils/BypassSlowDivision.cpp Tue Aug 21 10:15:33 2018
@@ -388,6 +388,15 @@ Optional<QuotRemPair> FastDivInsertionTa
return None;
}
+ // After Constant Hoisting pass, long constants may be represented as
+ // bitcast instructions. As a result, some constants may look like an
+ // instruction at first, and an additional check is necessary to find out if
+ // an operand is actually a constant.
+ if (auto *BCI = dyn_cast<BitCastInst>(Divisor))
+ if (BCI->getParent() == SlowDivOrRem->getParent() &&
+ isa<ConstantInt>(BCI->getOperand(0)))
+ return None;
+
if (DividendShort && !isSignedOp()) {
// If the division is unsigned and Dividend is known to be short, then
// either
Modified: llvm/trunk/test/CodeGen/X86/divide-by-constant.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/divide-by-constant.ll?rev=340303&r1=340302&r2=340303&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/divide-by-constant.ll (original)
+++ llvm/trunk/test/CodeGen/X86/divide-by-constant.ll Tue Aug 21 10:15:33 2018
@@ -384,39 +384,18 @@ define { i64, i32 } @PR38622(i64) nounwi
; X32-NEXT: popl %ebp
; X32-NEXT: retl
;
-; X64-FAST-LABEL: PR38622:
-; X64-FAST: # %bb.0:
-; X64-FAST-NEXT: movq %rdi, %rax
-; X64-FAST-NEXT: shrq $11, %rax
-; X64-FAST-NEXT: movabsq $4835703278458517, %rcx # imm = 0x112E0BE826D695
-; X64-FAST-NEXT: mulq %rcx
-; X64-FAST-NEXT: shrq $9, %rdx
-; X64-FAST-NEXT: imull $-294967296, %edx, %eax # imm = 0xEE6B2800
-; X64-FAST-NEXT: subl %eax, %edi
-; X64-FAST-NEXT: movq %rdx, %rax
-; X64-FAST-NEXT: movl %edi, %edx
-; X64-FAST-NEXT: retq
-;
-; X64-SLOW-LABEL: PR38622:
-; X64-SLOW: # %bb.0:
-; X64-SLOW-NEXT: movl $4000000000, %ecx # imm = 0xEE6B2800
-; X64-SLOW-NEXT: movq %rdi, %rax
-; X64-SLOW-NEXT: shrq $32, %rax
-; X64-SLOW-NEXT: je .LBB14_1
-; X64-SLOW-NEXT: # %bb.2:
-; X64-SLOW-NEXT: xorl %edx, %edx
-; X64-SLOW-NEXT: movq %rdi, %rax
-; X64-SLOW-NEXT: divq %rcx
-; X64-SLOW-NEXT: # kill: def $edx killed $edx killed $rdx
-; X64-SLOW-NEXT: retq
-; X64-SLOW-NEXT: .LBB14_1:
-; X64-SLOW-NEXT: xorl %edx, %edx
-; X64-SLOW-NEXT: movl %edi, %eax
-; X64-SLOW-NEXT: divl %ecx
-; X64-SLOW-NEXT: # kill: def $edx killed $edx def $rdx
-; X64-SLOW-NEXT: # kill: def $eax killed $eax def $rax
-; X64-SLOW-NEXT: # kill: def $edx killed $edx killed $rdx
-; X64-SLOW-NEXT: retq
+; X64-LABEL: PR38622:
+; X64: # %bb.0:
+; X64-NEXT: movq %rdi, %rax
+; X64-NEXT: shrq $11, %rax
+; X64-NEXT: movabsq $4835703278458517, %rcx # imm = 0x112E0BE826D695
+; X64-NEXT: mulq %rcx
+; X64-NEXT: shrq $9, %rdx
+; X64-NEXT: imull $-294967296, %edx, %eax # imm = 0xEE6B2800
+; X64-NEXT: subl %eax, %edi
+; X64-NEXT: movq %rdx, %rax
+; X64-NEXT: movl %edi, %edx
+; X64-NEXT: retq
%2 = udiv i64 %0, 4000000000
%3 = urem i64 %0, 4000000000
%4 = trunc i64 %3 to i32
@@ -459,41 +438,20 @@ define { i64, i32 } @PR38622_signed(i64)
; X32-NEXT: popl %ebp
; X32-NEXT: retl
;
-; X64-FAST-LABEL: PR38622_signed:
-; X64-FAST: # %bb.0:
-; X64-FAST-NEXT: movabsq $1237940039285380275, %rcx # imm = 0x112E0BE826D694B3
-; X64-FAST-NEXT: movq %rdi, %rax
-; X64-FAST-NEXT: imulq %rcx
-; X64-FAST-NEXT: movq %rdx, %rcx
-; X64-FAST-NEXT: shrq $63, %rcx
-; X64-FAST-NEXT: sarq $28, %rdx
-; X64-FAST-NEXT: leaq (%rdx,%rcx), %rax
-; X64-FAST-NEXT: addl %ecx, %edx
-; X64-FAST-NEXT: imull $-294967296, %edx, %ecx # imm = 0xEE6B2800
-; X64-FAST-NEXT: subl %ecx, %edi
-; X64-FAST-NEXT: movl %edi, %edx
-; X64-FAST-NEXT: retq
-;
-; X64-SLOW-LABEL: PR38622_signed:
-; X64-SLOW: # %bb.0:
-; X64-SLOW-NEXT: movl $4000000000, %ecx # imm = 0xEE6B2800
-; X64-SLOW-NEXT: movq %rdi, %rax
-; X64-SLOW-NEXT: shrq $32, %rax
-; X64-SLOW-NEXT: je .LBB15_1
-; X64-SLOW-NEXT: # %bb.2:
-; X64-SLOW-NEXT: movq %rdi, %rax
-; X64-SLOW-NEXT: cqto
-; X64-SLOW-NEXT: idivq %rcx
-; X64-SLOW-NEXT: # kill: def $edx killed $edx killed $rdx
-; X64-SLOW-NEXT: retq
-; X64-SLOW-NEXT: .LBB15_1:
-; X64-SLOW-NEXT: xorl %edx, %edx
-; X64-SLOW-NEXT: movl %edi, %eax
-; X64-SLOW-NEXT: divl %ecx
-; X64-SLOW-NEXT: # kill: def $edx killed $edx def $rdx
-; X64-SLOW-NEXT: # kill: def $eax killed $eax def $rax
-; X64-SLOW-NEXT: # kill: def $edx killed $edx killed $rdx
-; X64-SLOW-NEXT: retq
+; X64-LABEL: PR38622_signed:
+; X64: # %bb.0:
+; X64-NEXT: movabsq $1237940039285380275, %rcx # imm = 0x112E0BE826D694B3
+; X64-NEXT: movq %rdi, %rax
+; X64-NEXT: imulq %rcx
+; X64-NEXT: movq %rdx, %rcx
+; X64-NEXT: shrq $63, %rcx
+; X64-NEXT: sarq $28, %rdx
+; X64-NEXT: leaq (%rdx,%rcx), %rax
+; X64-NEXT: addl %ecx, %edx
+; X64-NEXT: imull $-294967296, %edx, %ecx # imm = 0xEE6B2800
+; X64-NEXT: subl %ecx, %edi
+; X64-NEXT: movl %edi, %edx
+; X64-NEXT: retq
%2 = sdiv i64 %0, 4000000000
%3 = srem i64 %0, 4000000000
%4 = trunc i64 %3 to i32
More information about the llvm-commits
mailing list