llvm-project/llvm/test/CodeGen/PowerPC/ec-input.ll

156 lines
10 KiB
LLVM
Raw Normal View History

[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
; RUN: llc -verify-machineinstrs < %s | FileCheck %s
; This test case used to fail both with and without -verify-machineinstrs
; (-verify-machineinstrs would catch the problem right after instruction
; scheduling because the live intervals would not be right for the registers
; that were both inputs to the inline asm and also early-clobber outputs).
target datalayout = "E-m:e-i64:64-n32:64"
target triple = "powerpc64-bgq-linux"
%struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713 = type { i32, i8*, i8*, i8*, i8*, i8*, i8*, i8*, i8*, i8*, i8*, i8*, %struct._IO_marker.118.8248.32638.195238.200116.211498.218002.221254.222880.224506.226132.240766.244018.245644.248896.260278.271660.281416.283042.302554.304180.325318.326944.344712*, %struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713*, i32, i32, i64, i16, i8, [1 x i8], i8*, i64, i8*, i8*, i8*, i8*, i64, i32, [20 x i8] }
%struct._IO_marker.118.8248.32638.195238.200116.211498.218002.221254.222880.224506.226132.240766.244018.245644.248896.260278.271660.281416.283042.302554.304180.325318.326944.344712 = type { %struct._IO_marker.118.8248.32638.195238.200116.211498.218002.221254.222880.224506.226132.240766.244018.245644.248896.260278.271660.281416.283042.302554.304180.325318.326944.344712*, %struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713*, i32 }
@.str236 = external unnamed_addr constant [121 x i8], align 1
@.str294 = external unnamed_addr constant [49 x i8], align 1
; Function Attrs: nounwind
declare void @fprintf(%struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713* nocapture, i8* nocapture readonly, ...) #0
; Function Attrs: inlinehint nounwind
define void @_ZN4PAMI6Device2MU15ResourceManager46calculatePerCoreMUResourcesBasedOnAvailabilityEv(i32 %inp32, i64 %inp64) #1 align 2 {
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
; CHECK-LABEL: @_ZN4PAMI6Device2MU15ResourceManager46calculatePerCoreMUResourcesBasedOnAvailabilityEv
; CHECK: sc
entry:
%numFreeResourcesInSubgroup = alloca i32, align 4
%0 = ptrtoint i32* %numFreeResourcesInSubgroup to i64
br label %for.cond2.preheader
for.cond2.preheader: ; preds = %if.end23.3, %entry
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %if.end23.3 ]
%group.098 = phi i32 [ 0, %entry ], [ %inc37, %if.end23.3 ]
%minFreeBatIdsPerCore.097 = phi i64 [ 32, %entry ], [ %numFreeBatIdsInGroup.0.minFreeBatIdsPerCore.0, %if.end23.3 ]
%minFreeRecFifosPerCore.096 = phi i64 [ 16, %entry ], [ %minFreeRecFifosPerCore.1, %if.end23.3 ]
%minFreeInjFifosPerCore.095 = phi i64 [ 32, %entry ], [ %numFreeInjFifosInGroup.0.minFreeInjFifosPerCore.0, %if.end23.3 ]
%cmp5 = icmp eq i32 %inp32, 0
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
br i1 %cmp5, label %if.end, label %if.then
if.then: ; preds = %if.end23.2, %if.end23.1, %if.end23, %for.cond2.preheader
unreachable
if.end: ; preds = %for.cond2.preheader
%1 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv = zext i32 %1 to i64
%2 = call { i64, i64, i64, i64 } asm sideeffect "sc", "=&{r0},=&{r3},=&{r4},=&{r5},{r0},{r3},{r4},{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{cr0},~{memory}"(i64 1034, i64 %indvars.iv, i64 %0, i64 %inp64) #2
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
%cmp10 = icmp eq i32 0, 0
br i1 %cmp10, label %if.end14, label %if.then11
if.then11: ; preds = %if.end.3, %if.end.2, %if.end.1, %if.end
unreachable
if.end14: ; preds = %if.end
%3 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%cmp19 = icmp eq i32 %inp32, 0
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
br i1 %cmp19, label %if.end23, label %if.then20
if.then20: ; preds = %if.end14.3, %if.end14.2, %if.end14.1, %if.end14
%conv4.i65.lcssa = phi i32 [ %inp32, %if.end14 ], [ 0, %if.end14.1 ], [ %conv4.i65.2, %if.end14.2 ], [ %conv4.i65.3, %if.end14.3 ]
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
call void (%struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713*, i8*, ...) @fprintf(%struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713* undef, i8* getelementptr inbounds ([121 x i8], [121 x i8]* @.str236, i64 0, i64 0), i32 signext 2503) #3
call void (%struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713*, i8*, ...) @fprintf(%struct._IO_FILE.119.8249.32639.195239.200117.211499.218003.221255.222881.224507.226133.240767.244019.245645.248897.260279.271661.281417.283043.302555.304181.325319.326945.344713* undef, i8* getelementptr inbounds ([49 x i8], [49 x i8]* @.str294, i64 0, i64 0), i32 signext %conv4.i65.lcssa) #3
unreachable
if.end23: ; preds = %if.end14
%conv15 = zext i32 %3 to i64
%4 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv24 = zext i32 %4 to i64
%5 = call { i64, i64, i64, i64 } asm sideeffect "sc", "=&{r0},=&{r3},=&{r4},=&{r5},{r0},{r3},{r4},{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{cr0},~{memory}"(i64 1033, i64 0, i64 %0, i64 %inp64) #2
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
%cmp5.1 = icmp eq i32 0, 0
br i1 %cmp5.1, label %if.end.1, label %if.then
for.end38: ; preds = %if.end23.3
ret void
if.end.1: ; preds = %if.end23
%6 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv.1 = zext i32 %6 to i64
%add.1 = add nuw nsw i64 %conv.1, %conv
%7 = call { i64, i64, i64, i64 } asm sideeffect "sc", "=&{r0},=&{r3},=&{r4},=&{r5},{r0},{r3},{r4},{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{cr0},~{memory}"(i64 1034, i64 0, i64 %0, i64 %inp64) #2
%cmp10.1 = icmp eq i32 %inp32, 0
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
br i1 %cmp10.1, label %if.end14.1, label %if.then11
if.end14.1: ; preds = %if.end.1
%8 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%cmp19.1 = icmp eq i32 0, 0
br i1 %cmp19.1, label %if.end23.1, label %if.then20
if.end23.1: ; preds = %if.end14.1
%conv15.1 = zext i32 %8 to i64
%add16.1 = add nuw nsw i64 %conv15.1, %conv15
%9 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv24.1 = zext i32 %9 to i64
%add25.1 = add nuw nsw i64 %conv24.1, %conv24
%cmp5.2 = icmp eq i32 %inp32, 0
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
br i1 %cmp5.2, label %if.end.2, label %if.then
if.end.2: ; preds = %if.end23.1
%10 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv.2 = zext i32 %10 to i64
%add.2 = add nuw nsw i64 %conv.2, %add.1
%11 = call { i64, i64, i64, i64 } asm sideeffect "sc", "=&{r0},=&{r3},=&{r4},=&{r5},{r0},{r3},{r4},{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{cr0},~{memory}"(i64 1034, i64 %inp64, i64 %0, i64 %inp64) #2
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
%cmp10.2 = icmp eq i32 0, 0
br i1 %cmp10.2, label %if.end14.2, label %if.then11
if.end14.2: ; preds = %if.end.2
%12 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%13 = call { i64, i64, i64, i64 } asm sideeffect "sc", "=&{r0},=&{r3},=&{r4},=&{r5},{r0},{r3},{r4},{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{cr0},~{memory}"(i64 1035, i64 %inp64, i64 %0, i64 0) #2
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
%asmresult1.i64.2 = extractvalue { i64, i64, i64, i64 } %13, 1
%conv4.i65.2 = trunc i64 %asmresult1.i64.2 to i32
%cmp19.2 = icmp eq i32 %conv4.i65.2, 0
br i1 %cmp19.2, label %if.end23.2, label %if.then20
if.end23.2: ; preds = %if.end14.2
%conv15.2 = zext i32 %12 to i64
%add16.2 = add nuw nsw i64 %conv15.2, %add16.1
%14 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv24.2 = zext i32 %14 to i64
%add25.2 = add nuw nsw i64 %conv24.2, %add25.1
%cmp5.3 = icmp eq i32 0, 0
br i1 %cmp5.3, label %if.end.3, label %if.then
if.end.3: ; preds = %if.end23.2
%15 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%conv.3 = zext i32 %15 to i64
%add.3 = add nuw nsw i64 %conv.3, %add.2
%cmp10.3 = icmp eq i32 %inp32, 0
[InlineAsm] Remove EarlyClobber on registers that are also inputs When an inline asm call has an output register marked as early-clobber, but that same register is also an input operand, what should we do? GCC accepts this, and is documented to accept this for read/write operands saying, "Furthermore, if the earlyclobber operand is also a read/write operand, then that operand is written only after it's used." For write-only operands, the situation seems less clear, but I have at least one existing codebase that assumes this will work, in part because it has syscall macros like this: ({ \ register uint64_t r0 __asm__ ("r0") = (__NR_ ## name); \ register uint64_t r3 __asm__ ("r3") = ((uint64_t) (arg0)); \ register uint64_t r4 __asm__ ("r4") = ((uint64_t) (arg1)); \ register uint64_t r5 __asm__ ("r5") = ((uint64_t) (arg2)); \ __asm__ __volatile__ \ ("sc" \ : "=&r"(r0),"=&r"(r3),"=&r"(r4),"=&r"(r5) \ : "0"(r0), "1"(r3), "2"(r4), "3"(r5) \ : "r6","r7","r8","r9","r10","r11","r12","cr0","memory"); \ r3; \ }) Furthermore, with register aliases and subregister relationships that only the backend knows about, rejecting this in the frontend seems like a difficult proposition (if we wanted to do so). However, keeping the early-clobber flag on the INLINEASM MI does not work for us, because it will cause the register's live interval to end to soon (so it will not appear defined to be used as an input). Fortunately, fixing this does not seem hard: When forming the INLINEASM MI, check to see if any of the early-clobber outputs are also inputs, and if so, remove the early-clobber flag. llvm-svn: 235283
2015-04-20 08:01:30 +08:00
br i1 %cmp10.3, label %if.end14.3, label %if.then11
if.end14.3: ; preds = %if.end.3
%16 = load i32, i32* %numFreeResourcesInSubgroup, align 4
%17 = call { i64, i64, i64, i64 } asm sideeffect "sc", "=&{r0},=&{r3},=&{r4},=&{r5},{r0},{r3},{r4},{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{cr0},~{memory}"(i64 1035, i64 0, i64 %0, i64 0) #2
%asmresult1.i64.3 = extractvalue { i64, i64, i64, i64 } %17, 1
%conv4.i65.3 = trunc i64 %asmresult1.i64.3 to i32
%cmp19.3 = icmp eq i32 %conv4.i65.3, 0
br i1 %cmp19.3, label %if.end23.3, label %if.then20
if.end23.3: ; preds = %if.end14.3
%conv15.3 = zext i32 %16 to i64
%add16.3 = add nuw nsw i64 %conv15.3, %add16.2
%add25.3 = add nuw nsw i64 0, %add25.2
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 4
%cmp27 = icmp ult i64 %add.3, %minFreeInjFifosPerCore.095
%numFreeInjFifosInGroup.0.minFreeInjFifosPerCore.0 = select i1 %cmp27, i64 %add.3, i64 %minFreeInjFifosPerCore.095
%cmp30 = icmp ult i64 %add16.3, %minFreeRecFifosPerCore.096
%minFreeRecFifosPerCore.1 = select i1 %cmp30, i64 %add16.3, i64 %minFreeRecFifosPerCore.096
%cmp33 = icmp ult i64 %add25.3, %minFreeBatIdsPerCore.097
%numFreeBatIdsInGroup.0.minFreeBatIdsPerCore.0 = select i1 %cmp33, i64 %add25.3, i64 %minFreeBatIdsPerCore.097
%inc37 = add nuw nsw i32 %group.098, 1
%cmp = icmp ult i32 %inc37, 16
br i1 %cmp, label %for.cond2.preheader, label %for.end38
}
attributes #0 = { nounwind "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "target-cpu"="a2q" }
attributes #1 = { inlinehint nounwind "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "target-cpu"="a2q" }
attributes #2 = { nounwind }
attributes #3 = { cold nounwind }