[llvm] r341047 - [NFC][CodeGen][SelectionDAG] Tests for X % C == 0 codegen improvement.
Roman Lebedev via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 30 02:32:21 PDT 2018
Author: lebedevri
Date: Thu Aug 30 02:32:21 2018
New Revision: 341047
URL: http://llvm.org/viewvc/llvm-project?rev=341047&view=rev
Log:
[NFC][CodeGen][SelectionDAG] Tests for X % C == 0 codegen improvement.
Hacker's Delight 10-17: when C is constant,
the result of X % C == 0 can be computed more cheaply
without actually calculating the remainder.
The motivation is discussed here:
https://bugs.llvm.org/show_bug.cgi?id=35479.
Patch by: hermord (Dmytro Shynkevych)!
For https://reviews.llvm.org/D50222
Added:
llvm/trunk/test/CodeGen/AArch64/urem-seteq-optsize.ll
llvm/trunk/test/CodeGen/AArch64/urem-seteq.ll
llvm/trunk/test/CodeGen/X86/urem-seteq-optsize.ll
llvm/trunk/test/CodeGen/X86/urem-seteq.ll
Added: llvm/trunk/test/CodeGen/AArch64/urem-seteq-optsize.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AArch64/urem-seteq-optsize.ll?rev=341047&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/AArch64/urem-seteq-optsize.ll (added)
+++ llvm/trunk/test/CodeGen/AArch64/urem-seteq-optsize.ll Thu Aug 30 02:32:21 2018
@@ -0,0 +1,41 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s
+
+; On AArch64, division in expensive. BuildRemEqFold should therefore run even
+; when optimizing for size. Only optimizing for minimum size retains a plain div.
+
+define i32 @test_minsize(i32 %X) optsize minsize nounwind readnone {
+; CHECK-LABEL: test_minsize:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, #5
+; CHECK-NEXT: udiv w8, w0, w8
+; CHECK-NEXT: add w8, w8, w8, lsl #2
+; CHECK-NEXT: mov w9, #-10
+; CHECK-NEXT: cmp w0, w8
+; CHECK-NEXT: mov w8, #42
+; CHECK-NEXT: csel w0, w8, w9, eq
+; CHECK-NEXT: ret
+ %rem = urem i32 %X, 5
+ %cmp = icmp eq i32 %rem, 0
+ %ret = select i1 %cmp, i32 42, i32 -10
+ ret i32 %ret
+}
+
+define i32 @test_optsize(i32 %X) optsize nounwind readnone {
+; CHECK-LABEL: test_optsize:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, #52429
+; CHECK-NEXT: movk w8, #52428, lsl #16
+; CHECK-NEXT: umull x8, w0, w8
+; CHECK-NEXT: lsr x8, x8, #34
+; CHECK-NEXT: add w8, w8, w8, lsl #2
+; CHECK-NEXT: mov w9, #-10
+; CHECK-NEXT: cmp w0, w8
+; CHECK-NEXT: mov w8, #42
+; CHECK-NEXT: csel w0, w8, w9, eq
+; CHECK-NEXT: ret
+ %rem = urem i32 %X, 5
+ %cmp = icmp eq i32 %rem, 0
+ %ret = select i1 %cmp, i32 42, i32 -10
+ ret i32 %ret
+}
Added: llvm/trunk/test/CodeGen/AArch64/urem-seteq.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AArch64/urem-seteq.ll?rev=341047&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/AArch64/urem-seteq.ll (added)
+++ llvm/trunk/test/CodeGen/AArch64/urem-seteq.ll Thu Aug 30 02:32:21 2018
@@ -0,0 +1,153 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s
+
+; This tests the BuildREMEqFold optimization with UREM, i32, odd divisor, SETEQ.
+; The corresponding pseudocode is:
+; Q <- [N * multInv(5, 2^32)] <=> [N * 0xCCCCCCCD] <=> [N * (-858993459)]
+; res <- [Q <= (2^32 - 1) / 5] <=> [Q <= 858993459] <=> [Q < 858993460]
+define i32 @test_urem_odd(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_odd:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, #52429
+; CHECK-NEXT: movk w8, #52428, lsl #16
+; CHECK-NEXT: umull x8, w0, w8
+; CHECK-NEXT: lsr x8, x8, #34
+; CHECK-NEXT: add w8, w8, w8, lsl #2
+; CHECK-NEXT: cmp w0, w8
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 5
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This is like test_urem_odd, except the divisor has bit 30 set.
+define i32 @test_urem_odd_bit30(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_odd_bit30:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, #-11
+; CHECK-NEXT: umull x8, w0, w8
+; CHECK-NEXT: mov w9, #3
+; CHECK-NEXT: lsr x8, x8, #62
+; CHECK-NEXT: movk w9, #16384, lsl #16
+; CHECK-NEXT: msub w8, w8, w9, w0
+; CHECK-NEXT: cmp w8, #0 // =0
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 1073741827
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This is like test_urem_odd, except the divisor has bit 31 set.
+define i32 @test_urem_odd_bit31(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_odd_bit31:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, w0
+; CHECK-NEXT: lsl x9, x8, #30
+; CHECK-NEXT: sub x8, x9, x8
+; CHECK-NEXT: lsr x8, x8, #61
+; CHECK-NEXT: mov w9, #-2147483645
+; CHECK-NEXT: msub w8, w8, w9, w0
+; CHECK-NEXT: cmp w8, #0 // =0
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 2147483651
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This tests the BuildREMEqFold optimization with UREM, i16, even divisor, SETNE.
+; In this case, D <=> 14 <=> 7 * 2^1, so D0 = 7 and K = 1.
+; The corresponding pseudocode is:
+; Q <- [N * multInv(D0, 2^16)] <=> [N * multInv(7, 2^16)] <=> [N * 28087]
+; Q <- [Q >>rot K] <=> [Q >>rot 1]
+; res <- ![Q <= (2^16 - 1) / 7] <=> ![Q <= 9362] <=> [Q > 9362]
+define i16 @test_urem_even(i16 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_even:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w10, #9363
+; CHECK-NEXT: ubfx w9, w0, #1, #15
+; CHECK-NEXT: movk w10, #37449, lsl #16
+; CHECK-NEXT: umull x9, w9, w10
+; CHECK-NEXT: and w8, w0, #0xffff
+; CHECK-NEXT: lsr x9, x9, #34
+; CHECK-NEXT: orr w10, wzr, #0xe
+; CHECK-NEXT: msub w8, w9, w10, w8
+; CHECK-NEXT: cmp w8, #0 // =0
+; CHECK-NEXT: cset w0, ne
+; CHECK-NEXT: ret
+ %urem = urem i16 %X, 14
+ %cmp = icmp ne i16 %urem, 0
+ %ret = zext i1 %cmp to i16
+ ret i16 %ret
+}
+
+; This is like test_urem_even, except the divisor has bit 30 set.
+define i32 @test_urem_even_bit30(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_even_bit30:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, #-415
+; CHECK-NEXT: umull x8, w0, w8
+; CHECK-NEXT: mov w9, #104
+; CHECK-NEXT: lsr x8, x8, #62
+; CHECK-NEXT: movk w9, #16384, lsl #16
+; CHECK-NEXT: msub w8, w8, w9, w0
+; CHECK-NEXT: cmp w8, #0 // =0
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 1073741928
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This is like test_urem_odd, except the divisor has bit 31 set.
+define i32 @test_urem_even_bit31(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_even_bit31:
+; CHECK: // %bb.0:
+; CHECK-NEXT: mov w8, #65435
+; CHECK-NEXT: movk w8, #32767, lsl #16
+; CHECK-NEXT: umull x8, w0, w8
+; CHECK-NEXT: mov w9, #102
+; CHECK-NEXT: lsr x8, x8, #62
+; CHECK-NEXT: movk w9, #32768, lsl #16
+; CHECK-NEXT: msub w8, w8, w9, w0
+; CHECK-NEXT: cmp w8, #0 // =0
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 2147483750
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; We should not proceed with this fold if the divisor is 1 or -1
+define i32 @test_urem_one(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_one:
+; CHECK: // %bb.0:
+; CHECK-NEXT: orr w0, wzr, #0x1
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 1
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; We can lower remainder of division by powers of two much better elsewhere;
+; also, BuildREMEqFold does not work when the only odd factor of the divisor is 1.
+; This ensures we don't touch powers of two.
+define i32 @test_urem_pow2(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_pow2:
+; CHECK: // %bb.0:
+; CHECK-NEXT: tst w0, #0xf
+; CHECK-NEXT: cset w0, eq
+; CHECK-NEXT: ret
+ %urem = urem i32 %X, 16
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
Added: llvm/trunk/test/CodeGen/X86/urem-seteq-optsize.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/urem-seteq-optsize.ll?rev=341047&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/urem-seteq-optsize.ll (added)
+++ llvm/trunk/test/CodeGen/X86/urem-seteq-optsize.ll Thu Aug 30 02:32:21 2018
@@ -0,0 +1,81 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=i686-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,X86
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,X64,NOBMI2
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,BMI2
+
+; On X86, division in expensive. BuildRemEqFold should therefore run even
+; when optimizing for size. Only optimizing for minimum size retains a plain div.
+
+define i32 @test_minsize(i32 %X) optsize minsize nounwind readnone {
+; X86-LABEL: test_minsize:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT: pushl $5
+; X86-NEXT: popl %ecx
+; X86-NEXT: xorl %edx, %edx
+; X86-NEXT: divl %ecx
+; X86-NEXT: testl %edx, %edx
+; X86-NEXT: je .LBB0_1
+; X86-NEXT: # %bb.2:
+; X86-NEXT: pushl $-10
+; X86-NEXT: popl %eax
+; X86-NEXT: retl
+; X86-NEXT: .LBB0_1:
+; X86-NEXT: pushl $42
+; X86-NEXT: popl %eax
+; X86-NEXT: retl
+;
+; X64-LABEL: test_minsize:
+; X64: # %bb.0:
+; X64-NEXT: pushq $5
+; X64-NEXT: popq %rcx
+; X64-NEXT: xorl %edx, %edx
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: divl %ecx
+; X64-NEXT: testl %edx, %edx
+; X64-NEXT: pushq $42
+; X64-NEXT: popq %rcx
+; X64-NEXT: pushq $-10
+; X64-NEXT: popq %rax
+; X64-NEXT: cmovel %ecx, %eax
+; X64-NEXT: retq
+ %rem = urem i32 %X, 5
+ %cmp = icmp eq i32 %rem, 0
+ %ret = select i1 %cmp, i32 42, i32 -10
+ ret i32 %ret
+}
+
+define i32 @test_optsize(i32 %X) optsize nounwind readnone {
+; X86-LABEL: test_optsize:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl $-858993459, %edx # imm = 0xCCCCCCCD
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: mull %edx
+; X86-NEXT: shrl $2, %edx
+; X86-NEXT: leal (%edx,%edx,4), %eax
+; X86-NEXT: cmpl %eax, %ecx
+; X86-NEXT: movl $42, %eax
+; X86-NEXT: je .LBB1_2
+; X86-NEXT: # %bb.1:
+; X86-NEXT: movl $-10, %eax
+; X86-NEXT: .LBB1_2:
+; X86-NEXT: retl
+;
+; X64-LABEL: test_optsize:
+; X64: # %bb.0:
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: movl $3435973837, %ecx # imm = 0xCCCCCCCD
+; X64-NEXT: imulq %rax, %rcx
+; X64-NEXT: shrq $34, %rcx
+; X64-NEXT: leal (%rcx,%rcx,4), %eax
+; X64-NEXT: cmpl %eax, %edi
+; X64-NEXT: movl $42, %ecx
+; X64-NEXT: movl $-10, %eax
+; X64-NEXT: cmovel %ecx, %eax
+; X64-NEXT: retq
+ %rem = urem i32 %X, 5
+ %cmp = icmp eq i32 %rem, 0
+ %ret = select i1 %cmp, i32 42, i32 -10
+ ret i32 %ret
+}
Added: llvm/trunk/test/CodeGen/X86/urem-seteq.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/urem-seteq.ll?rev=341047&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/urem-seteq.ll (added)
+++ llvm/trunk/test/CodeGen/X86/urem-seteq.ll Thu Aug 30 02:32:21 2018
@@ -0,0 +1,248 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=i686-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,X86
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,X64,NOBMI2
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,BMI2
+
+; This tests the BuildREMEqFold optimization with UREM, i32, odd divisor, SETEQ.
+; The corresponding pseudocode is:
+; Q <- [N * multInv(5, 2^32)] <=> [N * 0xCCCCCCCD] <=> [N * (-858993459)]
+; res <- [Q <= (2^32 - 1) / 5] <=> [Q <= 858993459] <=> [Q < 858993460]
+define i32 @test_urem_odd(i32 %X) nounwind readnone {
+; X86-LABEL: test_urem_odd:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl $-858993459, %edx # imm = 0xCCCCCCCD
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: mull %edx
+; X86-NEXT: shrl $2, %edx
+; X86-NEXT: leal (%edx,%edx,4), %edx
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: cmpl %edx, %ecx
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_odd:
+; X64: # %bb.0:
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: movl $3435973837, %ecx # imm = 0xCCCCCCCD
+; X64-NEXT: imulq %rax, %rcx
+; X64-NEXT: shrq $34, %rcx
+; X64-NEXT: leal (%rcx,%rcx,4), %ecx
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: cmpl %ecx, %edi
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %urem = urem i32 %X, 5
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This is like test_urem_odd, except the divisor has bit 30 set.
+define i32 @test_urem_odd_bit30(i32 %X) nounwind readnone {
+; X86-LABEL: test_urem_odd_bit30:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl $-11, %edx
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: mull %edx
+; X86-NEXT: shrl $30, %edx
+; X86-NEXT: imull $1073741827, %edx, %edx # imm = 0x40000003
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: cmpl %edx, %ecx
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_odd_bit30:
+; X64: # %bb.0:
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: movl $4294967285, %ecx # imm = 0xFFFFFFF5
+; X64-NEXT: imulq %rax, %rcx
+; X64-NEXT: shrq $62, %rcx
+; X64-NEXT: imull $1073741827, %ecx, %ecx # imm = 0x40000003
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: cmpl %ecx, %edi
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %urem = urem i32 %X, 1073741827
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This is like test_urem_odd, except the divisor has bit 31 set.
+define i32 @test_urem_odd_bit31(i32 %X) nounwind readnone {
+; X86-LABEL: test_urem_odd_bit31:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl $1073741823, %edx # imm = 0x3FFFFFFF
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: mull %edx
+; X86-NEXT: shrl $29, %edx
+; X86-NEXT: imull $-2147483645, %edx, %edx # imm = 0x80000003
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: cmpl %edx, %ecx
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_odd_bit31:
+; X64: # %bb.0:
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: movq %rax, %rcx
+; X64-NEXT: shlq $30, %rcx
+; X64-NEXT: subq %rax, %rcx
+; X64-NEXT: shrq $61, %rcx
+; X64-NEXT: imull $-2147483645, %ecx, %ecx # imm = 0x80000003
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: cmpl %ecx, %edi
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %urem = urem i32 %X, 2147483651
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This tests the BuildREMEqFold optimization with UREM, i16, even divisor, SETNE.
+; In this case, D <=> 14 <=> 7 * 2^1, so D0 = 7 and K = 1.
+; The corresponding pseudocode is:
+; Q <- [N * multInv(D0, 2^16)] <=> [N * multInv(7, 2^16)] <=> [N * 28087]
+; Q <- [Q >>rot K] <=> [Q >>rot 1]
+; res <- ![Q <= (2^16 - 1) / 7] <=> ![Q <= 9362] <=> [Q > 9362]
+define i16 @test_urem_even(i16 %X) nounwind readnone {
+; X86-LABEL: test_urem_even:
+; X86: # %bb.0:
+; X86-NEXT: movzwl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: shrl %eax
+; X86-NEXT: imull $18725, %eax, %eax # imm = 0x4925
+; X86-NEXT: shrl $17, %eax
+; X86-NEXT: movl %eax, %edx
+; X86-NEXT: shll $4, %edx
+; X86-NEXT: subl %eax, %edx
+; X86-NEXT: subl %eax, %edx
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: cmpw %dx, %cx
+; X86-NEXT: setne %al
+; X86-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_even:
+; X64: # %bb.0:
+; X64-NEXT: movzwl %di, %ecx
+; X64-NEXT: movl %ecx, %eax
+; X64-NEXT: shrl %eax
+; X64-NEXT: imull $18725, %eax, %eax # imm = 0x4925
+; X64-NEXT: shrl $17, %eax
+; X64-NEXT: movl %eax, %edx
+; X64-NEXT: shll $4, %edx
+; X64-NEXT: subl %eax, %edx
+; X64-NEXT: subl %eax, %edx
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: cmpw %dx, %cx
+; X64-NEXT: setne %al
+; X64-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-NEXT: retq
+ %urem = urem i16 %X, 14
+ %cmp = icmp ne i16 %urem, 0
+ %ret = zext i1 %cmp to i16
+ ret i16 %ret
+}
+
+; This is like test_urem_even, except the divisor has bit 30 set.
+define i32 @test_urem_even_bit30(i32 %X) nounwind readnone {
+; X86-LABEL: test_urem_even_bit30:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl $-415, %edx # imm = 0xFE61
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: mull %edx
+; X86-NEXT: shrl $30, %edx
+; X86-NEXT: imull $1073741928, %edx, %edx # imm = 0x40000068
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: cmpl %edx, %ecx
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_even_bit30:
+; X64: # %bb.0:
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: movl $4294966881, %ecx # imm = 0xFFFFFE61
+; X64-NEXT: imulq %rax, %rcx
+; X64-NEXT: shrq $62, %rcx
+; X64-NEXT: imull $1073741928, %ecx, %ecx # imm = 0x40000068
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: cmpl %ecx, %edi
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %urem = urem i32 %X, 1073741928
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; This is like test_urem_odd, except the divisor has bit 31 set.
+define i32 @test_urem_even_bit31(i32 %X) nounwind readnone {
+; X86-LABEL: test_urem_even_bit31:
+; X86: # %bb.0:
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: movl $2147483547, %edx # imm = 0x7FFFFF9B
+; X86-NEXT: movl %ecx, %eax
+; X86-NEXT: mull %edx
+; X86-NEXT: shrl $30, %edx
+; X86-NEXT: imull $-2147483546, %edx, %edx # imm = 0x80000066
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: cmpl %edx, %ecx
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_even_bit31:
+; X64: # %bb.0:
+; X64-NEXT: movl %edi, %eax
+; X64-NEXT: imulq $2147483547, %rax, %rax # imm = 0x7FFFFF9B
+; X64-NEXT: shrq $62, %rax
+; X64-NEXT: imull $-2147483546, %eax, %ecx # imm = 0x80000066
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: cmpl %ecx, %edi
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %urem = urem i32 %X, 2147483750
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; We should not proceed with this fold if the divisor is 1 or -1
+define i32 @test_urem_one(i32 %X) nounwind readnone {
+; CHECK-LABEL: test_urem_one:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movl $1, %eax
+; CHECK-NEXT: ret{{[l|q]}}
+ %urem = urem i32 %X, 1
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
+
+; We can lower remainder of division by powers of two much better elsewhere;
+; also, BuildREMEqFold does not work when the only odd factor of the divisor is 1.
+; This ensures we don't touch powers of two.
+define i32 @test_urem_pow2(i32 %X) nounwind readnone {
+; X86-LABEL: test_urem_pow2:
+; X86: # %bb.0:
+; X86-NEXT: xorl %eax, %eax
+; X86-NEXT: testb $15, {{[0-9]+}}(%esp)
+; X86-NEXT: sete %al
+; X86-NEXT: retl
+;
+; X64-LABEL: test_urem_pow2:
+; X64: # %bb.0:
+; X64-NEXT: xorl %eax, %eax
+; X64-NEXT: testb $15, %dil
+; X64-NEXT: sete %al
+; X64-NEXT: retq
+ %urem = urem i32 %X, 16
+ %cmp = icmp eq i32 %urem, 0
+ %ret = zext i1 %cmp to i32
+ ret i32 %ret
+}
More information about the llvm-commits
mailing list