[Bug target/90024] [7/8 Regression] ICE on AArch32 NEON mov with TImode constant.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90024 Matthew Malcomson changed: What|Removed |Added Target Milestone|7.5 |7.6
[Bug target/90024] [7/8 Regression] ICE on AArch32 NEON mov with TImode constant.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90024 Matthew Malcomson changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|--- |FIXED --- Comment #5 from Matthew Malcomson --- Fixed
[Bug target/90024] [7/8 Regression] ICE on AArch32 NEON mov with TImode constant.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90024 --- Comment #4 from Matthew Malcomson --- Author: matmal01 Date: Wed Apr 10 13:41:21 2019 New Revision: 270254 URL: https://gcc.gnu.org/viewcvs?rev=270254=gcc=rev Log: Backport of r270226 from mainline to gcc-8-branch The "*neon_mov" patterns for 128 bit sized quantities uses the "Dn" constraint to match vmov.f32 and vmov.i patterns. This constraint boils down to using the `neon_immediate_valid` function. Once the constraint has matched, the output C statement asserts that function passes. The output C statement calls `neon_immediate_valid` with the mode taken from the iterator, while the constraint takes the mode from the operand. This can cause a discrepency when the operand is a CONST_INT, as the constraint passes VOIDmode which `neon_immediate_valid` treats as DImode, while the C statement passes the mode of the iterator which can be TImode. When this happens, the `neon_immediate_valid` can fail in the second call (if e.g. the CONST_INT is a valid immediate in DImode but not TImode) which would trigger the assertion. The testcase added with this patch triggers this when compiled with an arm cross compiler using the command line below. gcc -march=armv8-a -c neon-immediate-timode.c -O1 -mfloat-abi=hard -mfpu=neon-fp-armv8 This patch splits the original "Dn" constraint into three new constraints, "DN" for TImode CONST_INT, "Dn" for DImode CONST_INT, and "Dm" for CONST_VECTOR. Splitting things up this way requires using one extra alternative in the "*neon_mov" patterns, but makes it clear from the constraint what mode is being used. We also remove the behaviour of treating VOIDmode as DImode in `neon_valid_immediate` since the original "Dn" constraint was the only place that functionality was used. VOIDmode is now never passed to that function. An assertion has been added to the function to ensure this problem is caught earlier on. bootstrapped and regtested on arm-none-linux-gnueabihf gcc/ChangeLog: 2019-04-10 Matthew Malcomson PR target/90024 * config/arm/arm.c (neon_valid_immediate): Disallow VOIDmode parameter. * config/arm/constraints.md (Dm, DN, Dn): Split previous Dn constraint into three. * config/arm/neon.md (*neon_mov): Account for TImode and DImode differences directly. (*smax3_neon, vashl3, vashr3_imm): Use Dm constraint. gcc/testsuite/ChangeLog: 2019-04-10 Matthew Malcomson PR target/90024 * gcc.dg/torture/neon-immediate-timode.c: New test. Added: branches/gcc-8-branch/gcc/testsuite/gcc.dg/torture/neon-immediate-timode.c Modified: branches/gcc-8-branch/gcc/ChangeLog branches/gcc-8-branch/gcc/config/arm/arm.c branches/gcc-8-branch/gcc/config/arm/constraints.md branches/gcc-8-branch/gcc/config/arm/neon.md branches/gcc-8-branch/gcc/testsuite/ChangeLog
[Bug target/90024] [7/8 Regression] ICE on AArch32 NEON mov with TImode constant.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90024 --- Comment #3 from Matthew Malcomson --- Author: matmal01 Date: Wed Apr 10 13:34:54 2019 New Revision: 270253 URL: https://gcc.gnu.org/viewcvs?rev=270253=gcc=rev Log: Backport of r270226 from mainline to gcc-7-branch The "*neon_mov" patterns for 128 bit sized quantities uses the "Dn" constraint to match vmov.f32 and vmov.i patterns. This constraint boils down to using the `neon_immediate_valid` function. Once the constraint has matched, the output C statement asserts that function passes. The output C statement calls `neon_immediate_valid` with the mode taken from the iterator, while the constraint takes the mode from the operand. This can cause a discrepency when the operand is a CONST_INT, as the constraint passes VOIDmode which `neon_immediate_valid` treats as DImode, while the C statement passes the mode of the iterator which can be TImode. When this happens, the `neon_immediate_valid` can fail in the second call (if e.g. the CONST_INT is a valid immediate in DImode but not TImode) which would trigger the assertion. The testcase added with this patch triggers this when compiled with an arm cross compiler using the command line below. gcc -march=armv8-a -c neon-immediate-timode.c -O1 -mfloat-abi=hard -mfpu=neon-fp-armv8 This patch splits the original "Dn" constraint into three new constraints, "DN" for TImode CONST_INT, "Dn" for DImode CONST_INT, and "Dm" for CONST_VECTOR. Splitting things up this way requires using one extra alternative in the "*neon_mov" patterns, but makes it clear from the constraint what mode is being used. We also remove the behaviour of treating VOIDmode as DImode in `neon_valid_immediate` since the original "Dn" constraint was the only place that functionality was used. VOIDmode is now never passed to that function. An assertion has been added to the function to ensure this problem is caught earlier on. bootstrapped and regtested on arm-none-linux-gnueabihf gcc/ChangeLog: 2019-04-10 Matthew Malcomson PR target/90024 * config/arm/arm.c (neon_valid_immediate): Disallow VOIDmode parameter. * config/arm/constraints.md (Dm, DN, Dn): Split previous Dn constraint into three. * config/arm/neon.md (*neon_mov): Account for TImode and DImode differences directly. (*smax3_neon, vashl3, vashr3_imm): Use Dm constraint. gcc/testsuite/ChangeLog: 2019-04-10 Matthew Malcomson PR target/90024 * gcc.dg/torture/neon-immediate-timode.c: New test. Added: branches/gcc-7-branch/gcc/testsuite/gcc.dg/torture/neon-immediate-timode.c Modified: branches/gcc-7-branch/gcc/ChangeLog branches/gcc-7-branch/gcc/config/arm/arm.c branches/gcc-7-branch/gcc/config/arm/constraints.md branches/gcc-7-branch/gcc/config/arm/neon.md branches/gcc-7-branch/gcc/testsuite/ChangeLog
[Bug target/90024] [7/8 Regression] ICE on AArch32 NEON mov with TImode constant.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90024 Richard Biener changed: What|Removed |Added Priority|P3 |P2 Known to work||9.0 Target Milestone|--- |7.5 Summary|[7/8/9 Regression] ICE on |[7/8 Regression] ICE on |AArch32 NEON mov with |AArch32 NEON mov with |TImode constant.|TImode constant.
